Daily Papers

We introduce an open source high-quality Mandarin TTS dataset MSceneSpeech (Multiple Scene Speech Dataset), which is intended to provide resources for expressive speech synthesis. MSceneSpeech comprises numerous audio recordings and texts performed and recorded according to daily life scenarios. Each scenario includes multiple speakers and a diverse range of prosodic styles, making it suitable for speech synthesis that entails multi-speaker style and prosody modeling. We have established a robust baseline, through the prompting mechanism, that can effectively synthesize speech characterized by both user-specific timbre and scene-specific prosody with arbitrary text input. The open source MSceneSpeech Dataset and audio samples of our baseline are available at https://speechai-demo.github.io/MSceneSpeech/.

The prosody of a spoken utterance, including features like stress, intonation and rhythm, can significantly affect the underlying semantics, and as a consequence can also affect its textual translation. Nevertheless, prosody is rarely studied within the context of speech-to-text translation (S2TT) systems. In particular, end-to-end (E2E) systems have been proposed as well-suited for prosody-aware translation because they have direct access to the speech signal when making translation decisions, but the understanding of whether this is successful in practice is still limited. A main challenge is the difficulty of evaluating prosody awareness in translation. To address this challenge, we introduce an evaluation methodology and a focused benchmark (named ContraProST) aimed at capturing a wide range of prosodic phenomena. Our methodology uses large language models and controllable text-to-speech (TTS) to generate contrastive examples. Through experiments in translating English speech into German, Spanish, and Japanese, we find that (a) S2TT models possess some internal representation of prosody, but the prosody signal is often not strong enough to affect the translations, (b) E2E systems outperform cascades of speech recognition and text translation systems, confirming their theoretical advantage in this regard, and (c) certain cascaded systems also capture prosodic information in the translation, but only to a lesser extent that depends on the particulars of the transcript's surface form.

Most End-to-End SLU methods depend on the pretrained ASR or language model features for intent prediction. However, other essential information in speech, such as prosody, is often ignored. Recent research has shown improved results in classifying dialogue acts by incorporating prosodic information. The margins of improvement in these methods are minimal as the neural models ignore prosodic features. In this work, we propose prosody-attention, which uses the prosodic features differently to generate attention maps across time frames of the utterance. Then we propose prosody-distillation to explicitly learn the prosodic information in the acoustic encoder rather than concatenating the implicit prosodic features. Both the proposed methods improve the baseline results, and the prosody-distillation method gives an intent classification accuracy improvement of 8\% and 2\% on SLURP and STOP datasets over the prosody baseline.

Non-verbal signals in speech are encoded by prosody and carry information that ranges from conversation action to attitude and emotion. Despite its importance, the principles that govern prosodic structure are not yet adequately understood. This paper offers an analytical schema and a technological proof-of-concept for the categorization of prosodic signals and their association with meaning. The schema interprets surface-representations of multi-layered prosodic events. As a first step towards implementation, we present a classification process that disentangles prosodic phenomena of three orders. It relies on fine-tuning a pre-trained speech recognition model, enabling the simultaneous multi-class/multi-label detection. It generalizes over a large variety of spontaneous data, performing on a par with, or superior to, human annotation. In addition to a standardized formalization of prosody, disentangling prosodic patterns can direct a theory of communication and speech organization. A welcome by-product is an interpretation of prosody that will enhance speech- and language-related technologies.

Prosody contains rich information beyond the literal meaning of words, which is crucial for the intelligibility of speech. Current models still fall short in phrasing and intonation; they not only miss or misplace breaks when synthesizing long sentences with complex structures but also produce unnatural intonation. We propose ProsodyFM, a prosody-aware text-to-speech synthesis (TTS) model with a flow-matching (FM) backbone that aims to enhance the phrasing and intonation aspects of prosody. ProsodyFM introduces two key components: a Phrase Break Encoder to capture initial phrase break locations, followed by a Duration Predictor for the flexible adjustment of break durations; and a Terminal Intonation Encoder which integrates a set of intonation shape tokens combined with a novel Pitch Processor for more robust modeling of human-perceived intonation change. ProsodyFM is trained with no explicit prosodic labels and yet can uncover a broad spectrum of break durations and intonation patterns. Experimental results demonstrate that ProsodyFM can effectively improve the phrasing and intonation aspects of prosody, thereby enhancing the overall intelligibility compared to four state-of-the-art (SOTA) models. Out-of-distribution experiments show that this prosody improvement can further bring ProsodyFM superior generalizability for unseen complex sentences and speakers. Our case study intuitively illustrates the powerful and fine-grained controllability of ProsodyFM over phrasing and intonation.

Speech sounds convey a great deal of information about the scenes, resulting in a variety of effects ranging from reverberation to additional ambient sounds. In this paper, we manipulate input speech to sound as though it was recorded within a different scene, given an audio-visual conditional example recorded from that scene. Our model learns through self-supervision, taking advantage of the fact that natural video contains recurring sound events and textures. We extract an audio clip from a video and apply speech enhancement. We then train a latent diffusion model to recover the original speech, using another audio-visual clip taken from elsewhere in the video as a conditional hint. Through this process, the model learns to transfer the conditional example's sound properties to the input speech. We show that our model can be successfully trained using unlabeled, in-the-wild videos, and that an additional visual signal can improve its sound prediction abilities. Please see our project webpage for video results: https://tinglok.netlify.app/files/avsoundscape/

We present a unified system to realize one-shot voice conversion (VC) on the pitch, rhythm, and speaker attributes. Existing works generally ignore the correlation between prosody and language content, leading to the degradation of naturalness in converted speech. Additionally, the lack of proper language features prevents these systems from accurately preserving language content after conversion. To address these issues, we devise a cascaded modular system leveraging self-supervised discrete speech units as language representation. These discrete units provide duration information essential for rhythm modeling. Our system first extracts utterance-level prosody and speaker representations from the raw waveform. Given the prosody representation, a prosody predictor estimates pitch, energy, and duration for each discrete unit in the utterance. A synthesizer further reconstructs speech based on the predicted prosody, speaker representation, and discrete units. Experiments show that our system outperforms previous approaches in naturalness, intelligibility, speaker transferability, and prosody transferability. Code and samples are publicly available.

Prosodic features such as pitch, timing, and intonation are central to spoken communication, conveying emotion, intent, and discourse structure. In text-based settings, where these cues are absent, emojis act as visual surrogates that add affective and pragmatic nuance. This study examines how emojis influence prosodic realisation in speech and how listeners interpret prosodic cues to recover emoji meanings. Unlike previous work, we directly link prosody and emoji by analysing actual human speech data, collected through structured but open-ended production and perception tasks. This provides empirical evidence of how emoji semantics shape spoken delivery and perception. Results show that speakers adapt their prosody based on emoji cues, listeners can often identify the intended emoji from prosodic variation alone, and greater semantic differences between emojis correspond to increased prosodic divergence. These findings suggest that emojis can act as meaningful carriers of prosodic intent, offering insight into their communicative role in digitally mediated contexts.

For a large portion of real-life utterances, the intention cannot be solely decided by either their semantic or syntactic characteristics. Although not all the sociolinguistic and pragmatic information can be digitized, at least phonetic features are indispensable in understanding the spoken language. Especially in head-final languages such as Korean, sentence-final prosody has great importance in identifying the speaker's intention. This paper suggests a system which identifies the inherent intention of a spoken utterance given its transcript, in some cases using auxiliary acoustic features. The main point here is a separate distinction for cases where discrimination of intention requires an acoustic cue. Thus, the proposed classification system decides whether the given utterance is a fragment, statement, question, command, or a rhetorical question/command, utilizing the intonation-dependency coming from the head-finality. Based on an intuitive understanding of the Korean language that is engaged in the data annotation, we construct a network which identifies the intention of a speech, and validate its utility with the test sentences. The system, if combined with up-to-date speech recognizers, is expected to be flexibly inserted into various language understanding modules.

From whirling ceiling fans to ticking clocks, the sounds that we hear subtly vary as we move through a scene. We ask whether these ambient sounds convey information about 3D scene structure and, if so, whether they provide a useful learning signal for multimodal models. To study this, we collect a dataset of paired audio and RGB-D recordings from a variety of quiet indoor scenes. We then train models that estimate the distance to nearby walls, given only audio as input. We also use these recordings to learn multimodal representations through self-supervision, by training a network to associate images with their corresponding sounds. These results suggest that ambient sound conveys a surprising amount of information about scene structure, and that it is a useful signal for learning multimodal features.

This paper presents Att-HACK, the first large database of acted speech with social attitudes. Available databases of expressive speech are rare and very often restricted to the primary emotions: anger, joy, sadness, fear. This greatly limits the scope of the research on expressive speech. Besides, a fundamental aspect of speech prosody is always ignored and missing from such databases: its variety, i.e. the possibility to repeat an utterance while varying its prosody. This paper represents a first attempt to widen the scope of expressivity in speech, by providing a database of acted speech with social attitudes: friendly, seductive, dominant, and distant. The proposed database comprises 25 speakers interpreting 100 utterances in 4 social attitudes, with 3-5 repetitions each per attitude for a total of around 30 hours of speech. The Att-HACK is freely available for academic research under a Creative Commons Licence.

Spontaneous speech has many affective and pragmatic functions that are interesting and challenging to model in TTS. However, the presence of reduced articulation, fillers, repetitions, and other disfluencies in spontaneous speech make the text and acoustics less aligned than in read speech, which is problematic for attention-based TTS. We propose a TTS architecture that can rapidly learn to speak from small and irregular datasets, while also reproducing the diversity of expressive phenomena present in spontaneous speech. Specifically, we add utterance-level prosody control to an existing neural HMM-based TTS system which is capable of stable, monotonic alignments for spontaneous speech. We objectively evaluate control accuracy and perform perceptual tests that demonstrate that prosody control does not degrade synthesis quality. To exemplify the power of combining prosody control and ecologically valid data for reproducing intricate spontaneous speech phenomena, we evaluate the system's capability of synthesizing two types of creaky voice. Audio samples are available at https://www.speech.kth.se/tts-demos/prosodic-hmm/

This paper investigates the application of voice activity projection (VAP), a predictive turn-taking model for spoken dialogue, on multilingual data, encompassing English, Mandarin, and Japanese. The VAP model continuously predicts the upcoming voice activities of participants in dyadic dialogue, leveraging a cross-attention Transformer to capture the dynamic interplay between participants. The results show that a monolingual VAP model trained on one language does not make good predictions when applied to other languages. However, a multilingual model, trained on all three languages, demonstrates predictive performance on par with monolingual models across all languages. Further analyses show that the multilingual model has learned to discern the language of the input signal. We also analyze the sensitivity to pitch, a prosodic cue that is thought to be important for turn-taking. Finally, we compare two different audio encoders, contrastive predictive coding (CPC) pre-trained on English, with a recent model based on multilingual wav2vec 2.0 (MMS).

Vocal feedback (e.g., `mhm', `yeah', `okay') is an important component of spoken dialogue and is crucial to ensuring common ground in conversational systems. The exact meaning of such feedback is conveyed through both lexical and prosodic form. In this work, we investigate the perceived prosodic similarity of vocal feedback with the same lexical form, and to what extent existing speech representations reflect such similarities. A triadic comparison task with recruited participants is used to measure perceived similarity of feedback responses taken from two different datasets. We find that spectral and self-supervised speech representations encode prosody better than extracted pitch features, especially in the case of feedback from the same speaker. We also find that it is possible to further condense and align the representations to human perception through contrastive learning.

Dubbing is a type of audiovisual translation where dialogues are translated and enacted so that they give the impression that the media is in the target language. It requires a careful alignment of dubbed recordings with the lip movements of performers in order to achieve visual coherence. In this paper, we deal with the specific problem of prosodic phrase synchronization within the framework of machine dubbing. Our methodology exploits the attention mechanism output in neural machine translation to find plausible phrasing for the translated dialogue lines and then uses them to condition their synthesis. Our initial work in this field records comparable speech rate ratio to professional dubbing translation, and improvement in terms of lip-syncing of long dialogue lines.

Speech language models refer to language models with speech processing and understanding capabilities. One key desirable capability for speech language models is the ability to capture the intricate interdependency between content and prosody. The existing mainstream paradigm of training speech language models, which converts speech into discrete tokens before feeding them into LLMs, is sub-optimal in learning prosody information -- we find that the resulting LLMs do not exhibit obvious emerging prosody processing capabilities via pre-training alone. To overcome this, we propose ProsodyLM, which introduces a simple tokenization scheme amenable to learning prosody. Each speech utterance is first transcribed into text, followed by a sequence of word-level prosody tokens. Compared with conventional speech tokenization schemes, the proposed tokenization scheme retains more complete prosody information, and is more understandable to text-based LLMs. We find that ProsodyLM can learn surprisingly diverse emerging prosody processing capabilities through pre-training alone, ranging from harnessing the prosody nuances in generated speech, such as contrastive focus, understanding emotion and stress in an utterance, to maintaining prosody consistency in long contexts.

Prosody plays an important role in characterizing the style of a speaker or an emotion, but most non-parallel voice or emotion style transfer algorithms do not convert any prosody information. Two major components of prosody are pitch and rhythm. Disentangling the prosody information, particularly the rhythm component, from the speech is challenging because it involves breaking the synchrony between the input speech and the disentangled speech representation. As a result, most existing prosody style transfer algorithms would need to rely on some form of text transcriptions to identify the content information, which confines their application to high-resource languages only. Recently, SpeechSplit has made sizeable progress towards unsupervised prosody style transfer, but it is unable to extract high-level global prosody style in an unsupervised manner. In this paper, we propose AutoPST, which can disentangle global prosody style from speech without relying on any text transcriptions. AutoPST is an Autoencoder-based Prosody Style Transfer framework with a thorough rhythm removal module guided by the self-expressive representation learning. Experiments on different style transfer tasks show that AutoPST can effectively convert prosody that correctly reflects the styles of the target domains.

Controlling text-to-speech (TTS) systems to synthesize speech with the prosodic characteristics expected by users has attracted much attention. To achieve controllability, current studies focus on two main directions: (1) using reference speech as prosody prompt to guide speech synthesis, and (2) using natural language descriptions to control the generation process. However, finding reference speech that exactly contains the prosody that users want to synthesize takes a lot of effort. Description-based guidance in TTS systems can only determine the overall prosody, which has difficulty in achieving fine-grained prosody control over the synthesized speech. In this paper, we propose DrawSpeech, a sketch-conditioned diffusion model capable of generating speech based on any prosody sketches drawn by users. Specifically, the prosody sketches are fed to DrawSpeech to provide a rough indication of the expected prosody trends. DrawSpeech then recovers the detailed pitch and energy contours based on the coarse sketches and synthesizes the desired speech. Experimental results show that DrawSpeech can generate speech with a wide variety of prosody and can precisely control the fine-grained prosody in a user-friendly manner. Our implementation and audio samples are publicly available.

In recent years, there has been growing interest in representing speech with discrete tokens, which serve as pseudo-text for speech language models (speechLMs) and as efficient intermediate representations for downstream tasks. These tokens are typically categorized as acoustic and phonetic tokens: the former holds detailed acoustic information for reconstruction while the latter mainly captures linguistic content. In human speech communication, however, unnecessary acoustic details such as speaker information are abstracted, while both linguistic and prosodic information are utilized for speech comprehension and production. Given this, neither type of token seems an ideal representation for tasks sensitive to prosody, such as speechLMs. In this study, we propose the Phonological Tokenizer, a method that fine-tunes phonetic tokens via differentiable k-means with a multi-task objective of ASR and speech resynthesis. Experimental validation on diverse tasks confirms that our tokens retain phonological (both linguistic and prosodic) information while appropriately discarding speaker identity.

Speech-to-speech translation systems today do not adequately support use for dialog purposes. In particular, nuances of speaker intent and stance can be lost due to improper prosody transfer. We present an exploration of what needs to be done to overcome this. First, we developed a data collection protocol in which bilingual speakers re-enact utterances from an earlier conversation in their other language, and used this to collect an English-Spanish corpus, so far comprising 1871 matched utterance pairs. Second, we developed a simple prosodic dissimilarity metric based on Euclidean distance over a broad set of prosodic features. We then used these to investigate cross-language prosodic differences, measure the likely utility of three simple baseline models, and identify phenomena which will require more powerful modeling. Our findings should inform future research on cross-language prosody and the design of speech-to-speech translation systems capable of effective prosody transfer.

In modern speech synthesis, paralinguistic information--such as a speaker's vocal timbre, emotional state, and dynamic prosody--plays a critical role in conveying nuance beyond mere semantics. Traditional Text-to-Speech (TTS) systems rely on fixed style labels or inserting a speech prompt to control these cues, which severely limits flexibility. Recent attempts seek to employ natural-language instructions to modulate paralinguistic features, substantially improving the generalization of instruction-driven TTS models. Although many TTS systems now support customized synthesis via textual description, their actual ability to interpret and execute complex instructions remains largely unexplored. In addition, there is still a shortage of high-quality benchmarks and automated evaluation metrics specifically designed for instruction-based TTS, which hinders accurate assessment and iterative optimization of these models. To address these limitations, we introduce InstructTTSEval, a benchmark for measuring the capability of complex natural-language style control. We introduce three tasks, namely Acoustic-Parameter Specification, Descriptive-Style Directive, and Role-Play, including English and Chinese subsets, each with 1k test cases (6k in total) paired with reference audio. We leverage Gemini as an automatic judge to assess their instruction-following abilities. Our evaluation of accessible instruction-following TTS systems highlights substantial room for further improvement. We anticipate that InstructTTSEval will drive progress toward more powerful, flexible, and accurate instruction-following TTS.

Recent research in zero-shot speech synthesis has made significant progress in speaker similarity. However, current efforts focus on timbre generalization rather than prosody modeling, which results in limited naturalness and expressiveness. To address this, we introduce a novel speech synthesis model trained on large-scale datasets, including both timbre and hierarchical prosody modeling. As timbre is a global attribute closely linked to expressiveness, we adopt a global vector to model speaker timbre while guiding prosody modeling. Besides, given that prosody contains both global consistency and local variations, we introduce a diffusion model as the pitch predictor and employ a prosody adaptor to model prosody hierarchically, further enhancing the prosody quality of the synthesized speech. Experimental results show that our model not only maintains comparable timbre quality to the baseline but also exhibits better naturalness and expressiveness.

Human speech exhibits rich and flexible prosodic variations. To address the one-to-many mapping problem from text to prosody in a reasonable and flexible manner, we propose DiffStyleTTS, a multi-speaker acoustic model based on a conditional diffusion module and an improved classifier-free guidance, which hierarchically models speech prosodic features, and controls different prosodic styles to guide prosody prediction. Experiments show that our method outperforms all baselines in naturalness and achieves superior synthesis speed compared to three diffusion-based baselines. Additionally, by adjusting the guiding scale, DiffStyleTTS effectively controls the guidance intensity of the synthetic prosody.

Improving text representation has attracted much attention to achieve expressive text-to-speech (TTS). However, existing works only implicitly learn the prosody with masked token reconstruction tasks, which leads to low training efficiency and difficulty in prosody modeling. We propose CLAPSpeech, a cross-modal contrastive pre-training framework that explicitly learns the prosody variance of the same text token under different contexts. Specifically, 1) We encourage the model to connect the text context with its corresponding prosody pattern in the joint multi-modal space with the elaborate design of the encoder inputs and contrastive loss; 2) We introduce a multi-scale pre-training pipeline to capture prosody patterns in multiple levels. We show how to incorporate CLAPSpeech into existing TTS models for better prosody. Experiments on three datasets not only show that CLAPSpeech could improve the prosody prediction for existing TTS methods, but also demonstrate its generalization ability to adapt to multiple languages and multi-speaker TTS. We also deeply analyze the principle behind the performance of CLAPSpeech. Ablation studies demonstrate the necessity of each component in our method. Source code and audio samples are available at https://clapspeech.github.io.

We present a multi-speaker Japanese audiobook text-to-speech (TTS) system that leverages multimodal context information of preceding acoustic context and bilateral textual context to improve the prosody of synthetic speech. Previous work either uses unilateral or single-modality context, which does not fully represent the context information. The proposed method uses an acoustic context encoder and a textual context encoder to aggregate context information and feeds it to the TTS model, which enables the model to predict context-dependent prosody. We conducted comprehensive objective and subjective evaluations on a multi-speaker Japanese audiobook dataset. Experimental results demonstrate that the proposed method significantly outperforms two previous works. Additionally, we present insights about the different choices of context - modalities, lateral information and length - for audiobook TTS that have never been discussed in the literature before.

Speech inherently contains rich acoustic information that extends far beyond the textual language. In real-world spoken language understanding, effective interpretation often requires integrating semantic meaning (e.g., content), paralinguistic features (e.g., emotions, speed, pitch) and phonological characteristics (e.g., prosody, intonation, rhythm), which are embedded in speech. While recent multimodal Speech Large Language Models (SpeechLLMs) have demonstrated remarkable capabilities in processing audio information, their ability to perform fine-grained perception and complex reasoning in natural speech remains largely unexplored. To address this gap, we introduce MMSU, a comprehensive benchmark designed specifically for understanding and reasoning in spoken language. MMSU comprises 5,000 meticulously curated audio-question-answer triplets across 47 distinct tasks. To ground our benchmark in linguistic theory, we systematically incorporate a wide range of linguistic phenomena, including phonetics, prosody, rhetoric, syntactics, semantics, and paralinguistics. Through a rigorous evaluation of 14 advanced SpeechLLMs, we identify substantial room for improvement in existing models, highlighting meaningful directions for future optimization. MMSU establishes a new standard for comprehensive assessment of spoken language understanding, providing valuable insights for developing more sophisticated human-AI speech interaction systems. MMSU benchmark is available at https://huggingface.co/datasets/ddwang2000/MMSU. Evaluation Code is available at https://github.com/dingdongwang/MMSU_Bench.

Multimodal generative models have shown impressive advances with the help of powerful diffusion models. Despite the progress, generating sound solely from text poses challenges in ensuring comprehensive scene depiction and temporal alignment. Meanwhile, video-to-sound generation limits the flexibility to prioritize sound synthesis for specific objects within the scene. To tackle these challenges, we propose a novel video-and-text-to-sound generation method, called ReWaS, where video serves as a conditional control for a text-to-audio generation model. Our method estimates the structural information of audio (namely, energy) from the video while receiving key content cues from a user prompt. We employ a well-performing text-to-sound model to consolidate the video control, which is much more efficient for training multimodal diffusion models with massive triplet-paired (audio-video-text) data. In addition, by separating the generative components of audio, it becomes a more flexible system that allows users to freely adjust the energy, surrounding environment, and primary sound source according to their preferences. Experimental results demonstrate that our method shows superiority in terms of quality, controllability, and training efficiency. Our demo is available at https://naver-ai.github.io/rewas

In recent years, audio generation has made significant progress in tasks such as text-to-speech (TTS), text-to-audio (TTA) and text-to-music (TTM). However, generating long-form and controllable audio from complex audio scene descriptions remains a significant challenge, as such scenes often require coordinated speech, sound effects, music, songs, temporal structure, and post-production. In this work, we introduce Audio-Oscar, a multi-agent framework for generating audio from complex descriptions. Audio-Oscar coordinates a set of specialist agents, each responsible for a different aspect of the audio scene, including character modeling and voice design, speech generation, fine-grained timeline planning, model selection, non-speech generation, and audio post-production. Audio-Oscar further incorporates feedback-driven refinement. In addition, to address the lack of suitable benchmarks for evaluating audio generation from complex audio scene descriptions, we construct ASG-Bench, an Audio Scene Generation Benchmark containing both scene descriptions paired with reference audio and text-only scene descriptions. Each scene is annotated with target audio events and temporal statements to evaluate whether the generated audio faithfully realizes the required scene content and temporal structure. Experimental results show that Audio-Oscar can effectively generate audio that matches complex scene descriptions. Project samples are available at https://audiooscar.github.io/. Our code is available at https://github.com/ziye26/Audio-Oscar.

Although humans engaged in face-to-face conversation simultaneously communicate both verbally and non-verbally, methods for joint and unified synthesis of speech audio and co-speech 3D gesture motion from text are a new and emerging field. These technologies hold great promise for more human-like, efficient, expressive, and robust synthetic communication, but are currently held back by the lack of suitably large datasets, as existing methods are trained on parallel data from all constituent modalities. Inspired by student-teacher methods, we propose a straightforward solution to the data shortage, by simply synthesising additional training material. Specifically, we use unimodal synthesis models trained on large datasets to create multimodal (but synthetic) parallel training data, and then pre-train a joint synthesis model on that material. In addition, we propose a new synthesis architecture that adds better and more controllable prosody modelling to the state-of-the-art method in the field. Our results confirm that pre-training on large amounts of synthetic data improves the quality of both the speech and the motion synthesised by the multimodal model, with the proposed architecture yielding further benefits when pre-trained on the synthetic data. See https://shivammehta25.github.io/MAGI/ for example output.

Interpretability research has shown that self-supervised Spoken Language Models (SLMs) encode a wide variety of features in human speech from the acoustic, phonetic, phonological, syntactic and semantic levels, to speaker characteristics. The bulk of prior research on representations of phonology has focused on segmental features such as phonemes; the encoding of suprasegmental phonology (such as tone and stress patterns) in SLMs is not yet well understood. Tone is a suprasegmental feature that is present in more than half of the world's languages. This paper aims to analyze the tone encoding capabilities of SLMs, using Mandarin and Vietnamese as case studies. We show that SLMs encode lexical tone to a significant degree even when they are trained on data from non-tonal languages. We further find that SLMs behave similarly to native and non-native human participants in tone and consonant perception studies, but they do not follow the same developmental trajectory.

Movie dubbing is the task of synthesizing speech from scripts conditioned on video scenes, requiring accurate lip sync, faithful timbre transfer, and proper modeling of character identity and emotion. However, existing methods face two major limitations: (1) high-quality multimodal dubbing datasets are limited in scale, suffer from high word error rates, contain sparse annotations, rely on costly manual labeling, and are restricted to monologue scenes, all of which hinder effective model training; (2) existing dubbing models rely solely on the lip region to learn audio-visual alignment, which limits their applicability to complex live-action cinematic scenes, and exhibit suboptimal performance in lip sync, speech quality, and emotional expressiveness. To address these issues, we propose FunCineForge, which comprises an end-to-end production pipeline for large-scale dubbing datasets and an MLLM-based dubbing model designed for diverse cinematic scenes. Using the pipeline, we construct the first Chinese television dubbing dataset with rich annotations, and demonstrate the high quality of these data. Experiments across monologue, narration, dialogue, and multi-speaker scenes show that our dubbing model consistently outperforms SOTA methods in audio quality, lip sync, timbre transfer, and instruction following. Code and demos are available at https://anonymous.4open.science/w/FunCineForge.

The sound effects that designers add to videos are designed to convey a particular artistic effect and, thus, may be quite different from a scene's true sound. Inspired by the challenges of creating a soundtrack for a video that differs from its true sound, but that nonetheless matches the actions occurring on screen, we propose the problem of conditional Foley. We present the following contributions to address this problem. First, we propose a pretext task for training our model to predict sound for an input video clip using a conditional audio-visual clip sampled from another time within the same source video. Second, we propose a model for generating a soundtrack for a silent input video, given a user-supplied example that specifies what the video should "sound like". We show through human studies and automated evaluation metrics that our model successfully generates sound from video, while varying its output according to the content of a supplied example. Project site: https://xypb.github.io/CondFoleyGen/

Aerial scene recognition is a fundamental task in remote sensing and has recently received increased interest. While the visual information from overhead images with powerful models and efficient algorithms yields considerable performance on scene recognition, it still suffers from the variation of ground objects, lighting conditions etc. Inspired by the multi-channel perception theory in cognition science, in this paper, for improving the performance on the aerial scene recognition, we explore a novel audiovisual aerial scene recognition task using both images and sounds as input. Based on an observation that some specific sound events are more likely to be heard at a given geographic location, we propose to exploit the knowledge from the sound events to improve the performance on the aerial scene recognition. For this purpose, we have constructed a new dataset named AuDio Visual Aerial sceNe reCognition datasEt (ADVANCE). With the help of this dataset, we evaluate three proposed approaches for transferring the sound event knowledge to the aerial scene recognition task in a multimodal learning framework, and show the benefit of exploiting the audio information for the aerial scene recognition. The source code is publicly available for reproducibility purposes.

Self-attention mechanisms have enabled transformers to achieve superhuman-level performance on many speech-to-text (STT) tasks, yet the challenge of automatic prosodic segmentation has remained unsolved. In this paper we finetune Whisper, a pretrained STT model, to annotate intonation unit (IU) boundaries by repurposing low-frequency tokens. Our approach achieves an accuracy of 95.8%, outperforming previous methods without the need for large-scale labeled data or enterprise grade compute resources. We also diminish input signals by applying a series of filters, finding that low pass filters at a 3.2 kHz level improve segmentation performance in out of sample and out of distribution contexts. We release our model as both a transcription tool and a baseline for further improvements in prosodic segmentation.

Accent normalization (AN) systems often struggle with unnatural outputs and undesired content distortion, stemming from both suboptimal training data and rigid duration modeling. In this paper, we propose a "source-synthesis" methodology for training data construction. By generating source L2 speech and using authentic native speech as the training target, our approach avoids learning from TTS artifacts and, crucially, requires no real L2 data in training. Alongside this data strategy, we introduce CosyAccent, a non-autoregressive model that resolves the trade-off between prosodic naturalness and duration control. CosyAccent implicitly models rhythm for flexibility yet offers explicit control over total output duration. Experiments show that, despite being trained without any real L2 speech, CosyAccent achieves significantly improved content preservation and superior naturalness compared to strong baselines trained on real-world data.

Audio and visual signals typically occur simultaneously, and humans possess an innate ability to correlate and synchronize information from these two modalities. Recently, a challenging problem known as Audio-Visual Segmentation (AVS) has emerged, intending to produce segmentation maps for sounding objects within a scene. However, the methods proposed so far have not sufficiently integrated audio and visual information, and the computational costs have been extremely high. Additionally, the outputs of different stages have not been fully utilized. To facilitate this research, we introduce a novel Progressive Confident Masking Attention Network (PMCANet). It leverages attention mechanisms to uncover the intrinsic correlations between audio signals and visual frames. Furthermore, we design an efficient and effective cross-attention module to enhance semantic perception by selecting query tokens. This selection is determined through confidence-driven units based on the network's multi-stage predictive outputs. Experiments demonstrate that our network outperforms other AVS methods while requiring less computational resources. The code is available at: https://github.com/PrettyPlate/PCMANet.

In this paper we introduce a new natural language processing dataset and benchmark for predicting prosodic prominence from written text. To our knowledge this will be the largest publicly available dataset with prosodic labels. We describe the dataset construction and the resulting benchmark dataset in detail and train a number of different models ranging from feature-based classifiers to neural network systems for the prediction of discretized prosodic prominence. We show that pre-trained contextualized word representations from BERT outperform the other models even with less than 10% of the training data. Finally we discuss the dataset in light of the results and point to future research and plans for further improving both the dataset and methods of predicting prosodic prominence from text. The dataset and the code for the models are publicly available.

Large language models (LLMs) trained purely on text ostensibly lack any direct perceptual experience, yet their internal representations are implicitly shaped by multimodal regularities encoded in language. We test the hypothesis that explicit sensory prompting can surface this latent structure, bringing a text-only LLM into closer representational alignment with specialist vision and audio encoders. When a sensory prompt tells the model to 'see' or 'hear', it cues the model to resolve its next-token predictions as if they were conditioned on latent visual or auditory evidence that is never actually supplied. Our findings reveal that lightweight prompt engineering can reliably activate modality-appropriate representations in purely text-trained LLMs.

We present RALL-E, a robust language modeling method for text-to-speech (TTS) synthesis. While previous work based on large language models (LLMs) shows impressive performance on zero-shot TTS, such methods often suffer from poor robustness, such as unstable prosody (weird pitch and rhythm/duration) and a high word error rate (WER), due to the autoregressive prediction style of language models. The core idea behind RALL-E is chain-of-thought (CoT) prompting, which decomposes the task into simpler steps to enhance the robustness of LLM-based TTS. To accomplish this idea, RALL-E first predicts prosody features (pitch and duration) of the input text and uses them as intermediate conditions to predict speech tokens in a CoT style. Second, RALL-E utilizes the predicted duration prompt to guide the computing of self-attention weights in Transformer to enforce the model to focus on the corresponding phonemes and prosody features when predicting speech tokens. Results of comprehensive objective and subjective evaluations demonstrate that, compared to a powerful baseline method VALL-E, RALL-E significantly improves the WER of zero-shot TTS from 6.3% (without reranking) and 2.1% (with reranking) to 2.8% and 1.0%, respectively. Furthermore, we demonstrate that RALL-E correctly synthesizes sentences that are hard for VALL-E and reduces the error rate from 68% to 4%.

This paper integrates graph-to-sequence into an end-to-end text-to-speech framework for syntax-aware modelling with syntactic information of input text. Specifically, the input text is parsed by a dependency parsing module to form a syntactic graph. The syntactic graph is then encoded by a graph encoder to extract the syntactic hidden information, which is concatenated with phoneme embedding and input to the alignment and flow-based decoding modules to generate the raw audio waveform. The model is experimented on two languages, English and Mandarin, using single-speaker, few samples of target speakers, and multi-speaker datasets, respectively. Experimental results show better prosodic consistency performance between input text and generated audio, and also get higher scores in the subjective prosodic evaluation, and show the ability of voice conversion. Besides, the efficiency of the model is largely boosted through the design of the AI chip operator with 5x acceleration.

Acoustic events are sounds with well-defined spectro-temporal characteristics which can be associated with the physical objects generating them. Acoustic scenes are collections of such acoustic events in no specific temporal order. Given this natural linkage between events and scenes, a common belief is that the ability to classify events must help in the classification of scenes. This has led to several efforts attempting to do well on Acoustic Event Tagging (AET) and Acoustic Scene Classification (ASC) using a multi-task network. However, in these efforts, improvement in one task does not guarantee an improvement in the other, suggesting a tension between ASC and AET. It is unclear if improvements in AET translates to improvements in ASC. We explore this conundrum through an extensive empirical study and show that under certain conditions, using AET as an auxiliary task in the multi-task network consistently improves ASC performance. Additionally, ASC performance further improves with the AET data-set size and is not sensitive to the choice of events or the number of events in the AET data-set. We conclude that this improvement in ASC performance comes from the regularization effect of using AET and not from the network's improved ability to discern between acoustic events.

Recent work reports gains in neural text-to-speech (TTS) with Group Relative Policy Optimization (GRPO). However, in the absence of a verifiable reward for prosody, GRPO trained on transcription-oriented signals (CER/NLL) lowers error rates yet collapses prosody into monotone, unnatural speech; adding speaker-similarity further destabilizes training and degrades CER. We address this with an iterative Direct Preference Optimization (DPO) scheme that uses only a few hundred human-labeled preference pairs per round to directly optimize prosodic naturalness while regularizing to the current model. On KoCC-TTS, a curated dataset of authentic Korean call center interactions capturing task-oriented dialogues, our method attains the highest human preference (ELO) with competitive CER, outperforming GRPO and strong commercial baselines. These results suggest that when prosody cannot be rewarded automatically, human preference optimization offers a practical and data-efficient path to natural and robust TTS. The demo page is available at https://tts.ch.dev

Speech representations from self-supervised speech models (S3Ms) are known to be sensitive to phonemic contrasts, but their sensitivity to prosodic contrasts has not been directly measured. The ABX discrimination task has been used to measure phonemic contrast in S3M representations via minimal pairs. We introduce prosodic ABX, an extension of this framework to evaluate prosodic contrast with only a handful of examples and no explicit labels. Also, we build and release a dataset of English and Japanese minimal pairs and use it along with a Mandarin dataset to evaluate contrast in English stress, Japanese pitch accent, and Mandarin tone. Finally, we show that model and layer rankings are often preserved across several experimental conditions, making it practical for low-resource settings.

Current strategies for achieving fine-grained prosody control in speech synthesis entail extracting additional style embeddings or adopting more complex architectures. To enable zero-shot application of pretrained text-to-speech (TTS) models, we present PRESENT (PRosody Editing without Style Embeddings or New Training), which exploits explicit prosody prediction in FastSpeech2-based models by modifying the inference process directly. We apply our text-to-prosody framework to zero-shot language transfer using a JETS model exclusively trained on English LJSpeech data. We obtain character error rates (CER) of 12.8%, 18.7% and 5.9% for German, Hungarian and Spanish respectively, beating the previous state-of-the-art CER by over 2x for all three languages. Furthermore, we allow subphoneme-level control, a first in this field. To evaluate its effectiveness, we show that PRESENT can improve the prosody of questions, and use it to generate Mandarin, a tonal language where vowel pitch varies at subphoneme level. We attain 25.3% hanzi CER and 13.0% pinyin CER with the JETS model. All our code and audio samples are available online.

Speech translation models are increasingly capable of preserving speech-specific information (e.g., speaker gender, prosody, and emphasis), yet evaluation metrics remain blind to such phenomena. We meta-evaluate both text- and speech-based quality estimation metrics on two contrastive datasets targeting gender agreement and prosody, and find that both fall short, even when given direct access to the speech signal. We then train SpeechCOMET, a family of quality estimation models with speech encoders, and evaluate a state-of-the-art SpeechLLM as a judge. Both match or exceed text-based COMET on standard quality estimation, but neither consistently assesses speech-specific phenomena. We identify three causes: (1) speech-specific features are not reliably preserved in current encoders, (2) models tend to ignore the speech source signal, and (3) quality estimation training data contains too few relevant examples. We release all models and code, and argue that progress requires dedicated speech-specific training data and models that genuinely condition on speech.

The thud of a bouncing ball, the onset of speech as lips open -- when visual and audio events occur together, it suggests that there might be a common, underlying event that produced both signals. In this paper, we argue that the visual and audio components of a video signal should be modeled jointly using a fused multisensory representation. We propose to learn such a representation in a self-supervised way, by training a neural network to predict whether video frames and audio are temporally aligned. We use this learned representation for three applications: (a) sound source localization, i.e. visualizing the source of sound in a video; (b) audio-visual action recognition; and (c) on/off-screen audio source separation, e.g. removing the off-screen translator's voice from a foreign official's speech. Code, models, and video results are available on our webpage: http://andrewowens.com/multisensory

Audio captioning aims to generate text descriptions from environmental sounds. One challenge of audio captioning is the difficulty of the generalization due to the lack of audio-text paired training data. In this work, we propose a simple yet effective method of dealing with small-scaled datasets by leveraging a pre-trained language model. We keep the language model frozen to maintain the expressivity for text generation, and we only learn to extract global and temporal features from the input audio. To bridge a modality gap between the audio features and the language model, we employ mapping networks that translate audio features to the continuous vectors the language model can understand, called prefixes. We evaluate our proposed method on the Clotho and AudioCaps dataset and show our method outperforms prior arts in diverse experimental settings.

Building text-to-speech (TTS) synthesisers for Indian languages is a difficult task owing to a large number of active languages. Indian languages can be classified into a finite set of families, prominent among them, Indo-Aryan and Dravidian. The proposed work exploits this property to build a generic TTS system using multiple languages from the same family in an end-to-end framework. Generic systems are quite robust as they are capable of capturing a variety of phonotactics across languages. These systems are then adapted to a new language in the same family using small amounts of adaptation data. Experiments indicate that good quality TTS systems can be built using only 7 minutes of adaptation data. An average degradation mean opinion score of 3.98 is obtained for the adapted TTSes. Extensive analysis of systematic interactions between languages in the generic TTSes is carried out. x-vectors are included as speaker embedding to synthesise text in a particular speaker's voice. An interesting observation is that the prosody of the target speaker's voice is preserved. These results are quite promising as they indicate the capability of generic TTSes to handle speaker and language switching seamlessly, along with the ease of adaptation to a new language.

Prosody diversity is essential for achieving naturalness and expressiveness in zero-shot text-to-speech (TTS). However, frequently used acoustic metrics capture only partial views of prosodic variation and correlate poorly with human perception, leaving the problem of reliably quantifying prosody diversity underexplored. To bridge this gap, we introduce ProsodyEval, a prosody diversity assessment dataset that provides Prosody Mean Opinion Score (PMOS) alongside conventional acoustic metrics. ProsodyEval comprises 1000 speech samples derived from 7 mainstream TTS systems, with 2000 human ratings. Building on this, we propose the Discretized Speech Weighted Edit Distance (DS-WED), a new objective diversity metric that quantifies prosodic variation via weighted edit distance over semantic tokens. Experiments on ProsodyEval show that DS-WED achieves substantially higher correlation with human judgments than existing acoustic metrics, while remaining highly robust in speech tokenization from HuBERT and WavLM. Leveraging DS-WED, we benchmark state-of-the-art open-source TTS systems on LibriSpeech test-clean and Seed-TTS test-en, and further explorations uncover several factors that influence prosody diversity, including generative modeling paradigms, duration control, and reinforcement learning. Moreover, we find that current large audio language models (LALMs) remain limited in capturing prosodic variations. Audio samples are available at https://prosodyeval.github.io.

Video-to-Speech (VTS) generation aims to synthesize speech from a silent video without auditory signals. However, existing VTS methods disregard the hierarchical nature of speech, which spans coarse speaker-aware semantics to fine-grained prosodic details. This oversight hinders direct alignment between visual and speech features at specific hierarchical levels during property matching. In this paper, leveraging the hierarchical structure of Residual Vector Quantization (RVQ)-based codec, we propose HiCoDiT, a novel Hierarchical Codec Diffusion Transformer that exploits the inherent hierarchy of discrete speech tokens to achieve strong audio-visual alignment. Specifically, since lower-level tokens encode coarse speaker-aware semantics and higher-level tokens capture fine-grained prosody, HiCoDiT employs low-level and high-level blocks to generate tokens at different levels. The low-level blocks condition on lip-synchronized motion and facial identity to capture speaker-aware content, while the high-level blocks use facial expression to modulate prosodic dynamics. Finally, to enable more effective coarse-to-fine conditioning, we propose a dual-scale adaptive instance layer normalization that jointly captures global vocal style through channel-wise normalization and local prosody dynamics through temporal-wise normalization. Extensive experiments demonstrate that HiCoDiT outperforms baselines in fidelity and expressiveness, highlighting the potential of discrete modelling for VTS. The code and speech demo are both available at https://github.com/Jiaxin-Ye/HiCoDiT.

Understanding emotion from speech requires sensitivity to both lexical and acoustic cues. However, it remains unclear whether large audio language models (LALMs) genuinely process acoustic information or rely primarily on lexical content. We present LISTEN (Lexical vs. Acoustic Speech Test for Emotion in Narratives), a controlled benchmark designed to disentangle lexical reliance from acoustic sensitivity in emotion understanding. Across evaluations of six state-of-the-art LALMs, we observe a consistent lexical dominance. Models predict "neutral" when lexical cues are neutral or absent, show limited gains under cue alignment, and fail to classify distinct emotions under cue conflict. In paralinguistic settings, performance approaches chance. These results indicate that current LALMs largely "transcribe" rather than "listen," relying heavily on lexical semantics while underutilizing acoustic cues. LISTEN offers a principled framework for assessing emotion understanding in multimodal models.

Recently, there has been an increasing focus on audio-text cross-modal learning. However, most of the existing audio-text datasets contain only simple descriptions of sound events. Compared with classification labels, the advantages of such descriptions are significantly limited. In this paper, we first analyze the detailed information that human descriptions of audio may contain beyond sound event labels. Based on the analysis, we propose an automatic pipeline for curating audio-text pairs with rich details. Leveraging the property that sounds can be mixed and concatenated in the time domain, we control details in four aspects: temporal relationship, loudness, speaker identity, and occurrence number, in simulating audio mixtures. Corresponding details are transformed into captions by large language models. Audio-text pairs with rich details in text descriptions are thereby obtained. We validate the effectiveness of our pipeline with a small amount of simulated data, demonstrating that the simulated data enables models to learn detailed audio captioning.

Given a piece of text, a video clip, and a reference audio, the movie dubbing task aims to generate speech that aligns with the video while cloning the desired voice. The existing methods have two primary deficiencies: (1) They struggle to simultaneously hold audio-visual sync and achieve clear pronunciation; (2) They lack the capacity to express user-defined emotions. To address these problems, we propose EmoDubber, an emotion-controllable dubbing architecture that allows users to specify emotion type and emotional intensity while satisfying high-quality lip sync and pronunciation. Specifically, we first design Lip-related Prosody Aligning (LPA), which focuses on learning the inherent consistency between lip motion and prosody variation by duration level contrastive learning to incorporate reasonable alignment. Then, we design Pronunciation Enhancing (PE) strategy to fuse the video-level phoneme sequences by efficient conformer to improve speech intelligibility. Next, the speaker identity adapting module aims to decode acoustics prior and inject the speaker style embedding. After that, the proposed Flow-based User Emotion Controlling (FUEC) is used to synthesize waveform by flow matching prediction network conditioned on acoustics prior. In this process, the FUEC determines the gradient direction and guidance scale based on the user's emotion instructions by the positive and negative guidance mechanism, which focuses on amplifying the desired emotion while suppressing others. Extensive experimental results on three benchmark datasets demonstrate favorable performance compared to several state-of-the-art methods.

Voice design from natural language aims to generate speaker timbres directly from free-form textual descriptions, allowing users to create voices tailored to specific roles, personalities, and emotions. Such controllable voice creation benefits a wide range of downstream applications-including storytelling, game dubbing, role-play agents, and conversational assistants, making it a significant task for modern Text-to-Speech models. However, existing models are largely trained on carefully recorded studio data, which produces speech that is clean and well-articulated, yet lacks the lived-in qualities of real human voices. To address these limitations, we present MOSS-VoiceGenerator, an open-source instruction-driven voice generation model that creates new timbres directly from natural language prompts. Motivated by the hypothesis that exposure to real-world acoustic variation produces more perceptually natural voices, we train on large-scale expressive speech data sourced from cinematic content. Subjective preference studies demonstrate its superiority in overall performance, instruction-following, and naturalness compared to other voice design models.

Recent Audio Multimodal Large Language Models (Audio MLLMs) demonstrate impressive performance on speech benchmarks, yet it remains unclear whether these models genuinely process acoustic signals or rely on text-based semantic inference. To systematically study this question, we introduce DEAF (Diagnostic Evaluation of Acoustic Faithfulness), a benchmark of over 2,700 conflict stimuli spanning three acoustic dimensions: emotional prosody, background sounds, and speaker identity. Then, we design a controlled multi-level evaluation framework that progressively increases textual influence, ranging from semantic conflicts in the content to misleading prompts and their combination, allowing us to disentangle content-driven bias from prompt-induced sycophancy. We further introduce diagnostic metrics to quantify model reliance on textual cues over acoustic signals. Our evaluation of seven Audio MLLMs reveals a consistent pattern of text dominance: models are sensitive to acoustic variations, yet predictions are predominantly driven by textual inputs, revealing a gap between high performance on standard speech benchmarks and genuine acoustic understanding.

Expressive text-to-speech systems have undergone significant advancements owing to prosody modeling, but conventional methods can still be improved. Traditional approaches have relied on the autoregressive method to predict the quantized prosody vector; however, it suffers from the issues of long-term dependency and slow inference. This study proposes a novel approach called DiffProsody in which expressive speech is synthesized using a diffusion-based latent prosody generator and prosody conditional adversarial training. Our findings confirm the effectiveness of our prosody generator in generating a prosody vector. Furthermore, our prosody conditional discriminator significantly improves the quality of the generated speech by accurately emulating prosody. We use denoising diffusion generative adversarial networks to improve the prosody generation speed. Consequently, DiffProsody is capable of generating prosody 16 times faster than the conventional diffusion model. The superior performance of our proposed method has been demonstrated via experiments.

Codec-based language models (LMs) have revolutionized text-to-speech (TTS). However, standard codecs entangle timbre and prosody, which hinders independent control in continuation-based LMs. To tackle this challenge, we propose DisCo-Speech, a zero-shot controllable TTS framework featuring a disentangled speech codec (DisCodec) and an LM-based generator. The core component DisCodec employs a two-stage design: 1) tri-factor disentanglement to separate speech into content, prosody, and timbre subspaces via parallel encoders and hybrid losses; and 2) fusion and reconstruction that merges content and prosody into unified content-prosody tokens suitable for LM prediction, while jointly optimizing reconstruction to address the disentanglement-reconstruction trade-off. This allows the LM to perform prosodic continuation from a style prompt while the decoder injects target timbre, enabling flexible zero-shot control. Experiments demonstrate that DisCo-Speech achieves competitive voice cloning and superior zero-shot prosody control. By resolving the core entanglement at the codec level, DisCo-Speech provides a robust foundation for controllable speech synthesis.

Paralinguistic vocalizations-including non-verbal sounds like laughter and breathing, as well as lexicalized interjections such as "uhm" and "oh"-are integral to natural spoken communication. Despite their importance in conveying affect, intent, and interactional cues, such cues remain largely overlooked in conventional automatic speech recognition (ASR) and text-to-speech (TTS) systems. We present NVSpeech, an integrated and scalable pipeline that bridges the recognition and synthesis of paralinguistic vocalizations, encompassing dataset construction, ASR modeling, and controllable TTS. (1) We introduce a manually annotated dataset of 48,430 human-spoken utterances with 18 word-level paralinguistic categories. (2) We develop the paralinguistic-aware ASR model, which treats paralinguistic cues as inline decodable tokens (e.g., "You're so funny [Laughter]"), enabling joint lexical and non-verbal transcription. This model is then used to automatically annotate a large corpus, the first large-scale Chinese dataset of 174,179 utterances (573 hours) with word-level alignment and paralingustic cues. (3) We finetune zero-shot TTS models on both human- and auto-labeled data to enable explicit control over paralinguistic vocalizations, allowing context-aware insertion at arbitrary token positions for human-like speech synthesis. By unifying the recognition and generation of paralinguistic vocalizations, NVSpeech offers the first open, large-scale, word-level annotated pipeline for expressive speech modeling in Mandarin, integrating recognition and synthesis in a scalable and controllable manner. Dataset and audio demos are available at https://nvspeech170k.github.io/.

Zero-shot text-to-speech aims at synthesizing voices with unseen speech prompts. Previous large-scale multispeaker TTS models have successfully achieved this goal with an enrolled recording within 10 seconds. However, most of them are designed to utilize only short speech prompts. The limited information in short speech prompts significantly hinders the performance of fine-grained identity imitation. In this paper, we introduce Mega-TTS 2, a generic zero-shot multispeaker TTS model that is capable of synthesizing speech for unseen speakers with arbitrary-length prompts. Specifically, we 1) design a multi-reference timbre encoder to extract timbre information from multiple reference speeches; 2) and train a prosody language model with arbitrary-length speech prompts; With these designs, our model is suitable for prompts of different lengths, which extends the upper bound of speech quality for zero-shot text-to-speech. Besides arbitrary-length prompts, we introduce arbitrary-source prompts, which leverages the probabilities derived from multiple P-LLM outputs to produce expressive and controlled prosody. Furthermore, we propose a phoneme-level auto-regressive duration model to introduce in-context learning capabilities to duration modeling. Experiments demonstrate that our method could not only synthesize identity-preserving speech with a short prompt of an unseen speaker but also achieve improved performance with longer speech prompts. Audio samples can be found in https://mega-tts.github.io/mega2_demo/.

Visual dubbing is the process of generating lip motions of an actor in a video to synchronise with given audio. Recent advances have made progress towards this goal but have not been able to produce an approach suitable for mass adoption. Existing methods are split into either person-generic or person-specific models. Person-specific models produce results almost indistinguishable from reality but rely on long training times using large single-person datasets. Person-generic works have allowed for the visual dubbing of any video to any audio without further training, but these fail to capture the person-specific nuances and often suffer from visual artefacts. Our method, based on data-efficient neural rendering priors, overcomes the limitations of existing approaches. Our pipeline consists of learning a deferred neural rendering prior network and actor-specific adaptation using neural textures. This method allows for high-quality visual dubbing with just a few seconds of data, that enables video dubbing for any actor - from A-list celebrities to background actors. We show that we achieve state-of-the-art in terms of visual quality and recognisability both quantitatively, and qualitatively through two user studies. Our prior learning and adaptation method generalises to limited data better and is more scalable than existing person-specific models. Our experiments on real-world, limited data scenarios find that our model is preferred over all others. The project page may be found at https://dubbingforeveryone.github.io/

What do deep neural speech models know about phonology? Existing work has examined the encoding of individual linguistic units such as phonemes in these models. Here we investigate interactions between units. Inspired by classic experiments on human speech perception, we study how Wav2Vec2 resolves phonotactic constraints. We synthesize sounds on an acoustic continuum between /l/ and /r/ and embed them in controlled contexts where only /l/, only /r/, or neither occur in English. Like humans, Wav2Vec2 models show a bias towards the phonotactically admissable category in processing such ambiguous sounds. Using simple measures to analyze model internals on the level of individual stimuli, we find that this bias emerges in early layers of the model's Transformer module. This effect is amplified by ASR finetuning but also present in fully self-supervised models. Our approach demonstrates how controlled stimulus designs can help localize specific linguistic knowledge in neural speech models.

Speech conveys not only linguistic information but also rich non-verbal vocal events such as laughing and crying. While semantic transcription is well-studied, the precise localization of non-verbal events remains a critical yet under-explored challenge. Current methods suffer from insufficient task definitions with limited category coverage and ambiguous temporal granularity. They also lack standardized evaluation frameworks, hindering the development of downstream applications. To bridge this gap, we first develop a refined taxonomy of 21 vocal events, with a new categorization into discrete (standalone) versus continuous (mixed with speech) types. Based on the refined taxonomy, we introduce WESR-Bench, an expert-annotated evaluation set (900+ utterances) with a novel position-aware protocol that disentangles ASR errors from event detection, enabling precise localization measurement for both discrete and continuous events. We also build a strong baseline by constructing a 1,700+ hour corpus, and train specialized models, surpassing both open-source audio-language models and commercial APIs while preserving ASR quality. We anticipate that WESR will serve as a foundational resource for future research in modeling rich, real-world auditory scenes.

Non-autoregressive text-to-speech (NAR-TTS) models such as FastSpeech 2 and Glow-TTS can synthesize high-quality speech from the given text in parallel. After analyzing two kinds of generative NAR-TTS models (VAE and normalizing flow), we find that: VAE is good at capturing the long-range semantics features (e.g., prosody) even with small model size but suffers from blurry and unnatural results; and normalizing flow is good at reconstructing the frequency bin-wise details but performs poorly when the number of model parameters is limited. Inspired by these observations, to generate diverse speech with natural details and rich prosody using a lightweight architecture, we propose PortaSpeech, a portable and high-quality generative text-to-speech model. Specifically, 1) to model both the prosody and mel-spectrogram details accurately, we adopt a lightweight VAE with an enhanced prior followed by a flow-based post-net with strong conditional inputs as the main architecture. 2) To further compress the model size and memory footprint, we introduce the grouped parameter sharing mechanism to the affine coupling layers in the post-net. 3) To improve the expressiveness of synthesized speech and reduce the dependency on accurate fine-grained alignment between text and speech, we propose a linguistic encoder with mixture alignment combining hard inter-word alignment and soft intra-word alignment, which explicitly extracts word-level semantic information. Experimental results show that PortaSpeech outperforms other TTS models in both voice quality and prosody modeling in terms of subjective and objective evaluation metrics, and shows only a slight performance degradation when reducing the model parameters to 6.7M (about 4x model size and 3x runtime memory compression ratio compared with FastSpeech 2). Our extensive ablation studies demonstrate that each design in PortaSpeech is effective.

Recent progress in Spoken Language Modeling has demonstrated the feasibility of learning language directly from speech. Generating speech through a pipeline that operates at the text level typically loses nuances, intonations, and non-verbal vocalizations. Modeling directly from speech opens up the path to more natural and expressive systems. On the other hand, speech-only systems tend to trail behind text-based language models in terms of their semantic abilities. We show that fine-tuning speech representation models on phoneme classification leads to more context-invariant representations, which in turn improve downstream language modeling performance.

Natural language is widely used to describe, prompt, and control audio systems, but rarely serves as the representation carrying audio itself. We introduce lexical acoustic coding (LAC), a framework in which pre-trained LLM sender and receiver agents transmit sound through natural language. Under fixed system prompts, the agents write their own analysis and synthesis code, communicating only through a lexical sentence, shared vocabulary, and optional symbolic music structure. The sender analyzes an input waveform into interpretable, non-learned acoustic descriptors, quantizes each with a feature-specific interval vocabulary, and verbalizes the lexical code as English. The receiver parses the sentence back into lexical-acoustic constraints and renders a waveform through closed-loop refinement. The transmitted text serves as both a rich caption and as the transport representation itself. We frame LAC as a finite-rate lossy quantizer, exposing trade-offs between vocabulary size, rate, and fidelity. Experiments on short sounds and symbolic music transfer show that plain text preserves measurable acoustic structure while remaining interpretable, editable, and native to LLM-mediated communication.

In this paper, we propose a novel framework to control voice style in prompt-based, controllable text-to-speech systems by leveraging textual personas as voice style prompts. We present two persona rewriting strategies to transform generic persona descriptions into speech-oriented prompts, enabling fine-grained manipulation of prosodic attributes such as pitch, emotion, and speaking rate. Experimental results demonstrate that our methods enhance the naturalness, clarity, and consistency of synthesized speech. Finally, we analyze implicit social biases introduced by LLM-based rewriting, with a focus on gender. We underscore voice style as a crucial factor for persona-driven AI dialogue systems.

Sound design involves creatively selecting, recording, and editing sound effects for various media like cinema, video games, and virtual/augmented reality. One of the most time-consuming steps when designing sound is synchronizing audio with video. In some cases, environmental recordings from video shoots are available, which can aid in the process. However, in video games and animations, no reference audio exists, requiring manual annotation of event timings from the video. We propose a system to extract repetitive actions onsets from a video, which are then used - in conjunction with audio or textual embeddings - to condition a diffusion model trained to generate a new synchronized sound effects audio track. In this way, we leave complete creative control to the sound designer while removing the burden of synchronization with video. Furthermore, editing the onset track or changing the conditioning embedding requires much less effort than editing the audio track itself, simplifying the sonification process. We provide sound examples, source code, and pretrained models to faciliate reproducibility

This paper describes the Microsoft end-to-end neural text to speech (TTS) system: DelightfulTTS for Blizzard Challenge 2021. The goal of this challenge is to synthesize natural and high-quality speech from text, and we approach this goal in two perspectives: The first is to directly model and generate waveform in 48 kHz sampling rate, which brings higher perception quality than previous systems with 16 kHz or 24 kHz sampling rate; The second is to model the variation information in speech through a systematic design, which improves the prosody and naturalness. Specifically, for 48 kHz modeling, we predict 16 kHz mel-spectrogram in acoustic model, and propose a vocoder called HiFiNet to directly generate 48 kHz waveform from predicted 16 kHz mel-spectrogram, which can better trade off training efficiency, modelling stability and voice quality. We model variation information systematically from both explicit (speaker ID, language ID, pitch and duration) and implicit (utterance-level and phoneme-level prosody) perspectives: 1) For speaker and language ID, we use lookup embedding in training and inference; 2) For pitch and duration, we extract the values from paired text-speech data in training and use two predictors to predict the values in inference; 3) For utterance-level and phoneme-level prosody, we use two reference encoders to extract the values in training, and use two separate predictors to predict the values in inference. Additionally, we introduce an improved Conformer block to better model the local and global dependency in acoustic model. For task SH1, DelightfulTTS achieves 4.17 mean score in MOS test and 4.35 in SMOS test, which indicates the effectiveness of our proposed system

This paper introduces a novel voice conversion (VC) model, guided by text instructions such as "articulate slowly with a deep tone" or "speak in a cheerful boyish voice". Unlike traditional methods that rely on reference utterances to determine the attributes of the converted speech, our model adds versatility and specificity to voice conversion. The proposed VC model is a neural codec language model which processes a sequence of discrete codes, resulting in the code sequence of converted speech. It utilizes text instructions as style prompts to modify the prosody and emotional information of the given speech. In contrast to previous approaches, which often rely on employing separate encoders like prosody and content encoders to handle different aspects of the source speech, our model handles various information of speech in an end-to-end manner. Experiments have demonstrated the impressive capabilities of our model in comprehending instructions and delivering reasonable results.

Standard text-to-speech (TTS) evaluation measures intelligibility (WER, CER) and overall naturalness (MOS, UTMOS) but does not quantify accent. A synthesiser may score well on all four yet sound non-native on features that are phonemic in the target language. For Indic languages, these features include retroflex articulation, aspiration, vowel length, and the Tamil retroflex approximant (letter zha). We present PSP, the Phoneme Substitution Profile, an interpretable, per-phonological-dimension accent benchmark for Indic TTS. PSP decomposes accent into six complementary dimensions: retroflex collapse rate (RR), aspiration fidelity (AF), vowel-length fidelity (LF), Tamil-zha fidelity (ZF), Frechet Audio Distance (FAD), and prosodic signature divergence (PSD). The first four are measured via forced alignment plus native-speaker-centroid acoustic probes over Wav2Vec2-XLS-R layer-9 embeddings; the latter two are corpus-level distributional distances. In this v1 we benchmark four commercial and open-source systems (ElevenLabs v3, Cartesia Sonic-3, Sarvam Bulbul, Indic Parler-TTS) on Hindi, Telugu, and Tamil pilot sets, with a fifth system (Praxy Voice) included on all three languages, plus an R5->R6 case study on Telugu. Three findings: (i) retroflex collapse grows monotonically with phonological difficulty Hindi < Telugu < Tamil (~1%, ~40%, ~68%); (ii) PSP ordering diverges from WER ordering -- commercial WER-leaders do not uniformly lead on retroflex or prosodic fidelity; (iii) no single system is Pareto-optimal across all six dimensions. We release native reference centroids (500 clips per language), 1000-clip embeddings for FAD, 500-clip prosodic feature matrices for PSD, 300-utterance golden sets per language, scoring code under MIT, and centroids under CC-BY. Formal MOS-correlation is deferred to v2; v1 reports five internal-consistency signals plus a native-audio sanity check.

Recent advances in text-to-audio (TTA) generation excel at synthesizing short audio clips but struggle with long-form narrative audio, which requires temporal coherence and compositional reasoning. To address this gap, we propose AudioStory, a unified framework that integrates large language models (LLMs) with TTA systems to generate structured, long-form audio narratives. AudioStory possesses strong instruction-following reasoning generation capabilities. It employs LLMs to decompose complex narrative queries into temporally ordered sub-tasks with contextual cues, enabling coherent scene transitions and emotional tone consistency. AudioStory has two appealing features: (1) Decoupled bridging mechanism: AudioStory disentangles LLM-diffuser collaboration into two specialized components, i.e., a bridging query for intra-event semantic alignment and a residual query for cross-event coherence preservation. (2) End-to-end training: By unifying instruction comprehension and audio generation within a single end-to-end framework, AudioStory eliminates the need for modular training pipelines while enhancing synergy between components. Furthermore, we establish a benchmark AudioStory-10K, encompassing diverse domains such as animated soundscapes and natural sound narratives. Extensive experiments show the superiority of AudioStory on both single-audio generation and narrative audio generation, surpassing prior TTA baselines in both instruction-following ability and audio fidelity. Our code is available at https://github.com/TencentARC/AudioStory

This paper explores the potential of constructing an AI spoken dialogue system that "thinks how to respond" and "thinks how to speak" simultaneously, which more closely aligns with the human speech production process compared to the current cascade pipeline of independent chatbot and Text-to-Speech (TTS) modules. We hypothesize that Large Language Models (LLMs) with billions of parameters possess significant speech understanding capabilities and can jointly model dialogue responses and linguistic features. We conduct two sets of experiments: 1) Prosodic structure prediction, a typical front-end task in TTS, demonstrating the speech understanding ability of LLMs, and 2) Further integrating dialogue response and a wide array of linguistic features using a unified encoding format. Our results indicate that the LLM-based approach is a promising direction for building unified spoken dialogue systems.

We propose PromptTTS++, a prompt-based text-to-speech (TTS) synthesis system that allows control over speaker identity using natural language descriptions. To control speaker identity within the prompt-based TTS framework, we introduce the concept of speaker prompt, which describes voice characteristics (e.g., gender-neutral, young, old, and muffled) designed to be approximately independent of speaking style. Since there is no large-scale dataset containing speaker prompts, we first construct a dataset based on the LibriTTS-R corpus with manually annotated speaker prompts. We then employ a diffusion-based acoustic model with mixture density networks to model diverse speaker factors in the training data. Unlike previous studies that rely on style prompts describing only a limited aspect of speaker individuality, such as pitch, speaking speed, and energy, our method utilizes an additional speaker prompt to effectively learn the mapping from natural language descriptions to the acoustic features of diverse speakers. Our subjective evaluation results show that the proposed method can better control speaker characteristics than the methods without the speaker prompt. Audio samples are available at https://reppy4620.github.io/demo.promptttspp/.

Representing wild sounds as images is an important but challenging task due to the lack of paired datasets between sound and images and the significant differences in the characteristics of these two modalities. Previous studies have focused on generating images from sound in limited categories or music. In this paper, we propose a novel approach to generate images from in-the-wild sounds. First, we convert sound into text using audio captioning. Second, we propose audio attention and sentence attention to represent the rich characteristics of sound and visualize the sound. Lastly, we propose a direct sound optimization with CLIPscore and AudioCLIP and generate images with a diffusion-based model. In experiments, it shows that our model is able to generate high quality images from wild sounds and outperforms baselines in both quantitative and qualitative evaluations on wild audio datasets.

This work introduces a new task, text-conditioned selective video-to-audio (V2A) generation, which produces only the user-intended sound from a multi-object video. This capability is especially crucial in multimedia production, where audio tracks are handled individually for each sound source for precise editing, mixing, and creative control. However, current approaches generate single source-mixed sounds at once, largely because visual features are entangled, and region cues or prompts often fail to specify the source. We propose SelVA, a novel text-conditioned V2A model that treats the text prompt as an explicit selector of target source and modulates video encoder to distinctly extract prompt-relevant video features. The proposed supplementary tokens promote cross-attention by suppressing text-irrelevant activations with efficient parameter tuning, yielding robust semantic and temporal grounding. SelVA further employs a self-augmentation scheme to overcome the lack of mono audio track supervision. We evaluate SelVA on VGG-MONOAUDIO, a curated benchmark of clean single-source videos for such a task. Extensive experiments and ablations consistently verify its effectiveness across audio quality, semantic alignment, and temporal synchronization. Code and demo are available at https://jnwnlee.github.io/selva-demo/.

The rapid evolution of end-to-end spoken dialogue systems demands transcending mere textual semantics to incorporate paralinguistic nuances and the spontaneous nature of human conversation. However, current methods struggle with two critical gaps: the modality gap, involving prosody and emotion, and the colloquialness gap, distinguishing written scripts from natural speech. To address these challenges, we introduce SDiaReward, an end-to-end multi-turn reward model trained on SDiaReward-Dataset, a novel collection of episode-level preference pairs explicitly targeting these gaps. It operates directly on full multi-turn speech episodes and is optimized with pairwise preference supervision, enabling joint assessment of modality and colloquialness in a single evaluator. We further establish ESDR-Bench, a stratified benchmark for robust episode-level evaluation. Experiments demonstrate that SDiaReward achieves state-of-the-art pairwise preference accuracy, significantly outperforming general-purpose audio LLMs. Further analysis suggests that SDiaReward captures relative conversational expressiveness beyond superficial synthesis cues, improving generalization across domains and recording conditions. Code, data, and demos are available at https://github.com/MM-Speech/SDiaReward/.

Advancements in spoken language processing have driven the development of spoken language models (SLMs), designed to achieve universal audio understanding by jointly learning text and audio representations for a wide range of tasks. Although promising results have been achieved, there is growing discussion regarding these models' generalization capabilities and the extent to which they truly integrate audio and text modalities in their internal representations. In this work, we evaluate four SLMs on the task of speech emotion recognition using a dataset of emotionally incongruent speech samples, a condition under which the semantic content of the spoken utterance conveys one emotion while speech expressiveness conveys another. Our results indicate that SLMs rely predominantly on textual semantics rather than speech emotion to perform the task, indicating that text-related representations largely dominate over acoustic representations. We release both the code and the Emotionally Incongruent Synthetic Speech dataset (EMIS) to the community.

Backchannels (e.g., `yeah', `mhm', and `right') are short, non-interruptive feedback signals whose lexical form and prosody jointly convey pragmatic meaning. While prior computational research has largely focused on predicting backchannel timing, the relationship between lexico-prosodic form and meaning remains underexplored. We propose a two-stage framework: first, fine-tuning large language models on dialogue transcripts to derive rich contextual representations; and second, learning a joint embedding space for dialogue contexts and backchannel realizations. We evaluate alignment with human perception via triadic similarity judgments (prosodic and cross-lexical) and a context-backchannel suitability task. Our results demonstrate that the learned projections substantially improve context-backchannel retrieval compared to previous methods. In addition, they reveal that backchannel form is highly sensitive to extended conversational context and that the learned embeddings align more closely with human judgments than raw WavLM features.

We present a speaker-aware approach for simulating multi-speaker conversations that captures temporal consistency and realistic turn-taking dynamics. Prior work typically models aggregate conversational statistics under an independence assumption across speakers and turns. In contrast, our method uses speaker-specific deviation distributions enforcing intra-speaker temporal consistency, while a Markov chain governs turn-taking and a fixed room impulse response preserves spatial realism. We also unify pauses and overlaps into a single gap distribution, modeled with kernel density estimation for smooth continuity. Evaluation on Switchboard using intrinsic metrics - global gap statistics, correlations between consecutive gaps, copula-based higher-order dependencies, turn-taking entropy, and gap survival functions - shows that speaker-aware simulation better aligns with real conversational patterns than the baseline method, capturing fine-grained temporal dependencies and realistic speaker alternation, while revealing open challenges in modeling long-range conversational structure.

Open-vocabulary audio-language models, like CLAP, offer a promising approach for zero-shot audio classification (ZSAC) by enabling classification with any arbitrary set of categories specified with natural language prompts. In this paper, we propose a simple but effective method to improve ZSAC with CLAP. Specifically, we shift from the conventional method of using prompts with abstract category labels (e.g., Sound of an organ) to prompts that describe sounds using their inherent descriptive features in a diverse context (e.g.,The organ's deep and resonant tones filled the cathedral.). To achieve this, we first propose ReCLAP, a CLAP model trained with rewritten audio captions for improved understanding of sounds in the wild. These rewritten captions describe each sound event in the original caption using their unique discriminative characteristics. ReCLAP outperforms all baselines on both multi-modal audio-text retrieval and ZSAC. Next, to improve zero-shot audio classification with ReCLAP, we propose prompt augmentation. In contrast to the traditional method of employing hand-written template prompts, we generate custom prompts for each unique label in the dataset. These custom prompts first describe the sound event in the label and then employ them in diverse scenes. Our proposed method improves ReCLAP's performance on ZSAC by 1%-18% and outperforms all baselines by 1% - 55%.

Language models increasingly serve as the backbone of text-to-speech (TTS) systems, yet we understand little about the representations they build when text and generated speech tokens share a single residual stream. We train BatchTopK sparse autoencoders on the LM backbone of CosyVoice3 and introduce a modality-aware auto-interp pipeline that labels each feature from where it fires-text-prefix context, 1-second speech clips, or both. The recovered features are interpretable, spanning phonemes, laughter, accent prompts and speaker gender. Steering through the SAE latent space shows these features are causal rather than merely descriptive: targeted interventions raise laughter probability from 0.02 to 0.79, flip perceived speaker gender, and control speech rate while preserving spoken content. SAE features thus serve both as interpretability objects and as control directions for TTS synthesis.

Humans naturally understand moments in a video by integrating visual and auditory cues. For example, localizing a scene in the video like "A scientist passionately speaks on wildlife conservation as dramatic orchestral music plays, with the audience nodding and applauding" requires simultaneous processing of visual, audio, and speech signals. However, existing models often struggle to effectively fuse and interpret audio information, limiting their capacity for comprehensive video temporal understanding. To address this, we present TriSense, a triple-modality large language model designed for holistic video temporal understanding through the integration of visual, audio, and speech modalities. Central to TriSense is a Query-Based Connector that adaptively reweights modality contributions based on the input query, enabling robust performance under modality dropout and allowing flexible combinations of available inputs. To support TriSense's multimodal capabilities, we introduce TriSense-2M, a high-quality dataset of over 2 million curated samples generated via an automated pipeline powered by fine-tuned LLMs. TriSense-2M includes long-form videos and diverse modality combinations, facilitating broad generalization. Extensive experiments across multiple benchmarks demonstrate the effectiveness of TriSense and its potential to advance multimodal video analysis. Code and dataset will be publicly released.

End-to-end models reach state-of-the-art performance for speech recognition, but global soft attention is not monotonic, which might lead to convergence problems, to instability, to bad generalisation, cannot be used for online streaming, and is also inefficient in calculation. Monotonicity can potentially fix all of this. There are several ad-hoc solutions or heuristics to introduce monotonicity, but a principled introduction is rarely found in literature so far. In this paper, we present a mathematically clean solution to introduce monotonicity, by introducing a new latent variable which represents the audio position or segment boundaries. We compare several monotonic latent models to our global soft attention baseline such as a hard attention model, a local windowed soft attention model, and a segmental soft attention model. We can show that our monotonic models perform as good as the global soft attention model. We perform our experiments on Switchboard 300h. We carefully outline the details of our training and release our code and configs.

Text-to-Speech (TTS) has recently seen great progress in synthesizing high-quality speech owing to the rapid development of parallel TTS systems, but producing speech with naturalistic prosodic variations, speaking styles and emotional tones remains challenging. Moreover, since duration and speech are generated separately, parallel TTS models still have problems finding the best monotonic alignments that are crucial for naturalistic speech synthesis. Here, we propose StyleTTS, a style-based generative model for parallel TTS that can synthesize diverse speech with natural prosody from a reference speech utterance. With novel Transferable Monotonic Aligner (TMA) and duration-invariant data augmentation schemes, our method significantly outperforms state-of-the-art models on both single and multi-speaker datasets in subjective tests of speech naturalness and speaker similarity. Through self-supervised learning of the speaking styles, our model can synthesize speech with the same prosodic and emotional tone as any given reference speech without the need for explicitly labeling these categories.

This paper presents VoiceLDM, a model designed to produce audio that accurately follows two distinct natural language text prompts: the description prompt and the content prompt. The former provides information about the overall environmental context of the audio, while the latter conveys the linguistic content. To achieve this, we adopt a text-to-audio (TTA) model based on latent diffusion models and extend its functionality to incorporate an additional content prompt as a conditional input. By utilizing pretrained contrastive language-audio pretraining (CLAP) and Whisper, VoiceLDM is trained on large amounts of real-world audio without manual annotations or transcriptions. Additionally, we employ dual classifier-free guidance to further enhance the controllability of VoiceLDM. Experimental results demonstrate that VoiceLDM is capable of generating plausible audio that aligns well with both input conditions, even surpassing the speech intelligibility of the ground truth audio on the AudioCaps test set. Furthermore, we explore the text-to-speech (TTS) and zero-shot text-to-audio capabilities of VoiceLDM and show that it achieves competitive results. Demos and code are available at https://voiceldm.github.io.

We often verbally express emotions in a multifaceted manner, they may vary in their intensities and may be expressed not just as a single but as a mixture of emotions. This wide spectrum of emotions is well-studied in the structural model of emotions, which represents variety of emotions as derivative products of primary emotions with varying degrees of intensity. In this paper, we propose an emotional text-to-speech design to simulate a wider spectrum of emotions grounded on the structural model. Our proposed design, Daisy-TTS, incorporates a prosody encoder to learn emotionally-separable prosody embedding as a proxy for emotion. This emotion representation allows the model to simulate: (1) Primary emotions, as learned from the training samples, (2) Secondary emotions, as a mixture of primary emotions, (3) Intensity-level, by scaling the emotion embedding, and (4) Emotions polarity, by negating the emotion embedding. Through a series of perceptual evaluations, Daisy-TTS demonstrated overall higher emotional speech naturalness and emotion perceiveability compared to the baseline.

Generative multimodal content is increasingly prevalent in much of the content creation arena, as it has the potential to allow artists and media personnel to create pre-production mockups by quickly bringing their ideas to life. The generation of audio from text prompts is an important aspect of such processes in the music and film industry. Many of the recent diffusion-based text-to-audio models focus on training increasingly sophisticated diffusion models on a large set of datasets of prompt-audio pairs. These models do not explicitly focus on the presence of concepts or events and their temporal ordering in the output audio with respect to the input prompt. Our hypothesis is focusing on how these aspects of audio generation could improve audio generation performance in the presence of limited data. As such, in this work, using an existing text-to-audio model Tango, we synthetically create a preference dataset where each prompt has a winner audio output and some loser audio outputs for the diffusion model to learn from. The loser outputs, in theory, have some concepts from the prompt missing or in an incorrect order. We fine-tune the publicly available Tango text-to-audio model using diffusion-DPO (direct preference optimization) loss on our preference dataset and show that it leads to improved audio output over Tango and AudioLDM2, in terms of both automatic- and manual-evaluation metrics.

Prior research indicates that LID model performance significantly declines on accented speech; however, the specific causes, extent, and characterization of these errors remain under-explored. (i) We identify a common failure mode on accented speech whereby LID systems often misclassify L2 accented speech as the speaker's native language or a related language. (ii) We present evidence suggesting that state-of-the-art models are invariant to permutations of short spans of speech, implying they classify on the basis of short phonotactic features indicative of accent rather than language. Our analysis reveals a simple method to enhance model robustness to accents through input chunking. (iii) We present an approach that integrates sequence-level information into our model without relying on monolingual ASR systems; this reduces accent-language confusion and significantly enhances performance on accented speech while maintaining comparable results on standard LID.

Audio chaptering, the task of automatically segmenting long-form audio into coherent sections, is increasingly important for navigating podcasts, lectures, and videos. Despite its relevance, research remains limited and text-based, leaving key questions unresolved about leveraging audio information, handling ASR errors, and transcript-free evaluation. We address these gaps through three contributions: (1) a systematic comparison between text-based models with acoustic features, a novel audio-only architecture (AudioSeg) operating on learned audio representations, and multimodal LLMs; (2) empirical analysis of factors affecting performance, including transcript quality, acoustic features, duration, and speaker composition; and (3) formalized evaluation protocols contrasting transcript-dependent text-space protocols with transcript-invariant time-space protocols. Our experiments on YTSeg reveal that AudioSeg substantially outperforms text-based approaches, pauses provide the largest acoustic gains, and MLLMs remain limited by context length and weak instruction following, yet MLLMs are promising on shorter audio.

Audio-video generation has often relied on complex multi-stage architectures or sequential synthesis of sound and visuals. We introduce Ovi, a unified paradigm for audio-video generation that models the two modalities as a single generative process. By using blockwise cross-modal fusion of twin-DiT modules, Ovi achieves natural synchronization and removes the need for separate pipelines or post hoc alignment. To facilitate fine-grained multimodal fusion modeling, we initialize an audio tower with an architecture identical to that of a strong pretrained video model. Trained from scratch on hundreds of thousands of hours of raw audio, the audio tower learns to generate realistic sound effects, as well as speech that conveys rich speaker identity and emotion. Fusion is obtained by jointly training the identical video and audio towers via blockwise exchange of timing (via scaled-RoPE embeddings) and semantics (through bidirectional cross-attention) on a vast video corpus. Our model enables cinematic storytelling with natural speech and accurate, context-matched sound effects, producing movie-grade video clips. All the demos, code and model weights are published at https://aaxwaz.github.io/Ovi

We study speech language models that incorporate semantic initialization and planning losses to achieve robust and consistent generation. Our approach initializes speech tokens with self-supervised features, applies a light alignment loss, and trains with thinning and auxiliary objectives that target robustness and content planning. We train three models: a 0.7B speech-only model, a 1.0B speech-only model, and a 1.0B interleaved model with both text and speech. Acoustic studies show that the speech-only models achieve the highest consistency across speaker, gender, sentiment, room, and background factors, surpassing larger systems. Interleaving improves lexical and syntactic probes and semantic--acoustic alignment but reduces consistency. Linear probes show that our initialization biases the model toward content structure while trading off prosody detail. These results show that LM-side design and training mix control the balance between acoustic stability and semantic grounding without changes to the tokenizer or runtime architecture. A demo and model weights are available for exploration.

Semantic information refers to the meaning conveyed through words, phrases, and contextual relationships within a given linguistic structure. Humans can leverage semantic information, such as familiar linguistic patterns and contextual cues, to reconstruct incomplete or masked speech signals in noisy environments. However, existing speech enhancement (SE) approaches often overlook the rich semantic information embedded in speech, which is crucial for improving intelligibility, speaker consistency, and overall quality of enhanced speech signals. To enrich the SE model with semantic information, we employ language models as an efficient semantic learner and propose a comprehensive framework tailored for language model-based speech enhancement, called GenSE. Specifically, we approach SE as a conditional language modeling task rather than a continuous signal regression problem defined in existing works. This is achieved by tokenizing speech signals into semantic tokens using a pre-trained self-supervised model and into acoustic tokens using a custom-designed single-quantizer neural codec model. To improve the stability of language model predictions, we propose a hierarchical modeling method that decouples the generation of clean semantic tokens and clean acoustic tokens into two distinct stages. Moreover, we introduce a token chain prompting mechanism during the acoustic token generation stage to ensure timbre consistency throughout the speech enhancement process. Experimental results on benchmark datasets demonstrate that our proposed approach outperforms state-of-the-art SE systems in terms of speech quality and generalization capability.

Generating a voice from a face image is crucial for developing virtual humans capable of interacting using their unique voices, without relying on pre-recorded human speech. In this paper, we propose Face-StyleSpeech, a zero-shot Text-To-Speech (TTS) synthesis model that generates natural speech conditioned on a face image rather than reference speech. We hypothesize that learning both speaker identity and prosody from a face image poses a significant challenge. To address the issue, our TTS model incorporates both a face encoder and a prosody encoder. The prosody encoder is specifically designed to model prosodic features that are not captured only with a face image, allowing the face encoder to focus solely on capturing the speaker identity from the face image. Experimental results demonstrate that Face-StyleSpeech effectively generates more natural speech from a face image than baselines, even for the face images the model has not trained. Samples are at our demo page https://face-stylespeech.github.io.

Recent advancements in visual speech recognition (VSR) have promoted progress in lip-to-speech synthesis, where pre-trained VSR models enhance the intelligibility of synthesized speech by providing valuable semantic information. The success achieved by cascade frameworks, which combine pseudo-VSR with pseudo-text-to-speech (TTS) or implicitly utilize the transcribed text, highlights the benefits of leveraging VSR models. However, these methods typically rely on mel-spectrograms as an intermediate representation, which may introduce a key bottleneck: the domain gap between synthetic mel-spectrograms, generated from inherently error-prone lip-to-speech mappings, and real mel-spectrograms used to train vocoders. This mismatch inevitably degrades synthesis quality. To bridge this gap, we propose Natural Lip-to-Speech (NaturalL2S), an end-to-end framework integrating acoustic inductive biases with differentiable speech generation components. Specifically, we introduce a fundamental frequency (F0) predictor to capture prosodic variations in synthesized speech. The predicted F0 then drives a Differentiable Digital Signal Processing (DDSP) synthesizer to generate a coarse signal which serves as prior information for subsequent speech synthesis. Additionally, instead of relying on a reference speaker embedding as an auxiliary input, our approach achieves satisfactory performance on speaker similarity without explicitly modelling speaker characteristics. Both objective and subjective evaluation results demonstrate that NaturalL2S can effectively enhance the quality of the synthesized speech when compared to state-of-the-art methods. Our demonstration page is accessible at https://yifan-liang.github.io/NaturalL2S/.

Automatic detection of phoneme or word-like units is one of the core objectives in zero-resource speech processing. Recent attempts employ self-supervised training methods, such as contrastive predictive coding (CPC), where the next frame is predicted given past context. However, CPC only looks at the audio signal's frame-level structure. We overcome this limitation with a segmental contrastive predictive coding (SCPC) framework that can model the signal structure at a higher level e.g. at the phoneme level. In this framework, a convolutional neural network learns frame-level representation from the raw waveform via noise-contrastive estimation (NCE). A differentiable boundary detector finds variable-length segments, which are then used to optimize a segment encoder via NCE to learn segment representations. The differentiable boundary detector allows us to train frame-level and segment-level encoders jointly. Typically, phoneme and word segmentation are treated as separate tasks. We unify them and experimentally show that our single model outperforms existing phoneme and word segmentation methods on TIMIT and Buckeye datasets. We analyze the impact of boundary threshold and when is the right time to include the segmental loss in the learning process.

Machine-generated speech is characterized by its limited or unnatural emotional variation. Current text to speech systems generates speech with either a flat emotion, emotion selected from a predefined set, average variation learned from prosody sequences in training data or transferred from a source style. We propose a text to speech(TTS) system, where a user can choose the emotion of generated speech from a continuous and meaningful emotion space (Arousal-Valence space). The proposed TTS system can generate speech from the text in any speaker's style, with fine control of emotion. We show that the system works on emotion unseen during training and can scale to previously unseen speakers given his/her speech sample. Our work expands the horizon of the state-of-the-art FastSpeech2 backbone to a multi-speaker setting and gives it much-coveted continuous (and interpretable) affective control, without any observable degradation in the quality of the synthesized speech.

The success of large language models in text processing has inspired their adaptation to speech modeling. However, since speech is continuous and complex, it is often discretized for autoregressive modeling. Speech tokens derived from self-supervised models (known as semantic tokens) typically focus on the linguistic aspects of speech but neglect prosodic information. As a result, models trained on these tokens can generate speech with reduced naturalness. Existing approaches try to fix this by adding pitch features to the semantic tokens. However, pitch alone cannot fully represent the range of paralinguistic attributes, and selecting the right features requires careful hand-engineering. To overcome this, we propose an end-to-end variational approach that automatically learns to encode these continuous speech attributes to enhance the semantic tokens. Our approach eliminates the need for manual extraction and selection of paralinguistic features. Moreover, it produces preferred speech continuations according to human raters. Code, samples and models are available at https://github.com/b04901014/vae-gslm.

Recent progress in large-scale zero-shot speech synthesis has been significantly advanced by language models and diffusion models. However, the generation process of both methods is slow and computationally intensive. Efficient speech synthesis using a lower computing budget to achieve quality on par with previous work remains a significant challenge. In this paper, we present FlashSpeech, a large-scale zero-shot speech synthesis system with approximately 5\% of the inference time compared with previous work. FlashSpeech is built on the latent consistency model and applies a novel adversarial consistency training approach that can train from scratch without the need for a pre-trained diffusion model as the teacher. Furthermore, a new prosody generator module enhances the diversity of prosody, making the rhythm of the speech sound more natural. The generation processes of FlashSpeech can be achieved efficiently with one or two sampling steps while maintaining high audio quality and high similarity to the audio prompt for zero-shot speech generation. Our experimental results demonstrate the superior performance of FlashSpeech. Notably, FlashSpeech can be about 20 times faster than other zero-shot speech synthesis systems while maintaining comparable performance in terms of voice quality and similarity. Furthermore, FlashSpeech demonstrates its versatility by efficiently performing tasks like voice conversion, speech editing, and diverse speech sampling. Audio samples can be found in https://flashspeech.github.io/.

Omnimodal large language models (OmniLLMs) have recently gained increasing attention for unified audio-video understanding. However, processing long multimodal token sequences introduces substantial computational overhead, making efficient token compression crucial. Existing methods typically rely on fixed, modality-specific guidance, which fails to account for the varying importance of modalities across different queries. To address this limitation, we propose OmniSelect, a training-free, modality-adaptive token pruning framework that dynamically selects appropriate compression strategies for multimodal inputs. Specifically, we leverage a lightweight AudioCLIP model to estimate cross-modal relevance and categorize each input into three pruning regimes: Audio-Centric, Video-Centric, and Uniform pruning. Based on these relevance scores, OmniSelect further performs fine-grained token pruning within each temporal group, adaptively allocating pruning ratios to preserve informative tokens across modalities. By explicitly modeling modality preference and enabling dynamic strategy selection, OmniSelect effectively avoids the pitfalls of one-size-fits-all compression. Extensive experiments demonstrate that our method achieves efficient multimodal token reduction while maintaining strong performance, without requiring any additional training.

Generative models for speech synthesis face a fundamental trade-off: discrete tokens ensure stability but sacrifice expressivity, while continuous signals retain acoustic richness but suffer from error accumulation due to task entanglement. This challenge has driven the field towards multi-stage pipelines that rely on pre-trained speech tokenizers, but these create a semantic-acoustic divide, limiting holistic and expressive speech generation. We resolve these dilemma through hierarchical semantic-acoustic modeling with semi-discrete residual representations and present a novel tokenizer-free TTS model VoxCPM. Our framework introduces a differentiable quantization bottleneck that induces natural specialization: a Text-Semantic Language Model (TSLM) generates semantic-prosodic plans, while a Residual Acoustic Model (RALM) recovers fine-grained acoustic details. This hierarchical semantic-acoustic representation guides a local diffusion-based decoder to generate high-fidelity speech latents. Critically, the entire architecture is trained end-to-end under a simple diffusion objective, eliminating dependency on external speech tokenizers. Trained on a massive 1.8 million hours of bilingual corpus, our VoxCPM-0.5B model achieves state-of-the-art zero-shot TTS performance among open-source systems, demonstrating that our approach delivers expressive and stable synthesis. Besides, VoxCPM shows the capability to comprehend text to infer and generate appropriate prosody and style, delivering speech with context-aware expressiveness and natural flow. To facilitate community-driven research and development, VoxCPM is publicly accessible under Apache 2.0.

Audio-driven video generation aims to synthesize realistic videos that align with input audio recordings, akin to the human ability to visualize scenes from auditory input. However, existing approaches predominantly focus on exploring semantic information, such as the classes of sounding sources present in the audio, limiting their ability to generate videos with accurate content and spatial composition. In contrast, we humans can not only naturally identify the semantic categories of sounding sources but also determine their deeply encoded spatial attributes, including locations and movement directions. This useful information can be elucidated by considering specific spatial indicators derived from the inherent physical properties of sound, such as loudness or frequency. As prior methods largely ignore this factor, we present SpA2V, the first framework explicitly exploits these spatial auditory cues from audios to generate videos with high semantic and spatial correspondence. SpA2V decomposes the generation process into two stages: 1) Audio-guided Video Planning: We meticulously adapt a state-of-the-art MLLM for a novel task of harnessing spatial and semantic cues from input audio to construct Video Scene Layouts (VSLs). This serves as an intermediate representation to bridge the gap between the audio and video modalities. 2) Layout-grounded Video Generation: We develop an efficient and effective approach to seamlessly integrate VSLs as conditional guidance into pre-trained diffusion models, enabling VSL-grounded video generation in a training-free manner. Extensive experiments demonstrate that SpA2V excels in generating realistic videos with semantic and spatial alignment to the input audios.

While recent work shows promising results in expanding the capabilities of large language models (LLM) to directly understand and synthesize speech, an LLM-based strategy for modeling spoken dialogs remains elusive and calls for further investigation. This work proposes an extensive speech-text LLM framework, named the Unified Spoken Dialog Model (USDM), to generate coherent spoken responses with organic prosodic features relevant to the given input speech without relying on automatic speech recognition (ASR) or text-to-speech (TTS) solutions. Our approach employs a multi-step speech-text inference scheme that leverages chain-of-reasoning capabilities exhibited by the underlying LLM. We also propose a generalized speech-text pretraining scheme that helps with capturing cross-modal semantics. Automatic and human evaluations show that the proposed approach is effective in generating natural-sounding spoken responses, outperforming both prior and cascaded baselines. Detailed comparative studies reveal that, despite the cascaded approach being stronger in individual components, the joint speech-text modeling improves robustness against recognition errors and speech quality. Demo is available at https://unifiedsdm.github.io.

Zero-shot Text-to-Speech (TTS) aims to synthesize high-quality speech that mimics the voice of an unseen speaker using only a short reference sample, requiring not only speaker adaptation but also accurate modeling of prosodic attributes. Recent approaches based on language models, diffusion, and flow matching have shown promising results in zero-shot TTS, but still suffer from slow inference and repetition artifacts. Discrete codec representations have been widely adopted for speech synthesis, and recent works have begun to explore diffusion models in purely discrete settings, suggesting the potential of discrete generative modeling for speech synthesis. However, existing flow-matching methods typically embed these discrete tokens into a continuous space and apply continuous flow matching, which may not fully leverage the advantages of discrete representations. To address these challenges, we introduce DiFlow-TTS, which, to the best of our knowledge, is the first model to explore purely Discrete Flow Matching for speech synthesis. DiFlow-TTS explicitly models factorized speech attributes within a compact and unified architecture. It leverages in-context learning by conditioning on textual content, along with prosodic and acoustic attributes extracted from a reference speech, enabling effective attribute cloning in a zero-shot setting. In addition, the model employs a factorized flow prediction mechanism with distinct heads for prosody and acoustic details, allowing it to learn aspect-specific distributions. Experimental results demonstrate that DiFlow-TTS achieves promising performance in several key metrics, including naturalness, prosody, preservation of speaker style, and energy control. It also maintains a compact model size and achieves low-latency inference, generating speech up to 25.8 times faster than the latest existing baselines.

Segment Anything Model 2 (SAM2) exhibits strong generalisation for promptable segmentation in video clips; however, its integration with the audio modality remains underexplored. Existing approaches either convert audio into visual prompts (e.g., boxes) via foundation models, or inject adapters into the image encoder for audio-visual fusion. Yet both directions fall short in human-in-the-loop scenarios due to limited prompt accuracy and increased inference overhead. In particular, these adapter-based methods often suffer from audio prompt dilution, where the signal gradually weakens as it propagates through the network. In this work, we propose AuralSAM2, which integrates audio into SAM2 while largely preserving its promptable segmentation capability. Its core module, AuralFuser, fuses audio and visual features to generate sparse and dense prompts. Guided by audio and built upon SAM2's feature pyramid, these prompts propagate auditory cues across visual layers, reinforcing cross-modal influence. To further align modalities, we introduce an audio-guided contrastive loss that emphasises auditory relevance in dominant visual features. Our method achieves notable accuracy gains on public benchmarks with only minimal impact on the interactive efficiency of promptable segmentation. Our code is available at https://github.com/yyliu01/AuralSAM2.

Generating accurate sounds for complex audio-visual scenes is challenging, especially in the presence of multiple objects and sound sources. In this paper, we propose an {\em interactive object-aware audio generation} model that grounds sound generation in user-selected visual objects within images. Our method integrates object-centric learning into a conditional latent diffusion model, which learns to associate image regions with their corresponding sounds through multi-modal attention. At test time, our model employs image segmentation to allow users to interactively generate sounds at the {\em object} level. We theoretically validate that our attention mechanism functionally approximates test-time segmentation masks, ensuring the generated audio aligns with selected objects. Quantitative and qualitative evaluations show that our model outperforms baselines, achieving better alignment between objects and their associated sounds. Project page: https://tinglok.netlify.app/files/avobject/

With read-aloud speech synthesis achieving high naturalness scores, there is a growing research interest in synthesising spontaneous speech. However, human spontaneous face-to-face conversation has both spoken and non-verbal aspects (here, co-speech gestures). Only recently has research begun to explore the benefits of jointly synthesising these two modalities in a single system. The previous state of the art used non-probabilistic methods, which fail to capture the variability of human speech and motion, and risk producing oversmoothing artefacts and sub-optimal synthesis quality. We present the first diffusion-based probabilistic model, called Diff-TTSG, that jointly learns to synthesise speech and gestures together. Our method can be trained on small datasets from scratch. Furthermore, we describe a set of careful uni- and multi-modal subjective tests for evaluating integrated speech and gesture synthesis systems, and use them to validate our proposed approach. Please see https://shivammehta25.github.io/Diff-TTSG/ for video examples, data, and code.

Existing large audio-language models perceive the world as "mono"-a single stream of audio that ignores the critical spatial dimension ("where") required for universal audio scene analysis (ASA). To bridge this gap, we first introduce a hierarchical framework for audio scene analysis. Guided by this framework, we introduce a system that enables large audio-language models (LALMs) to understand and reason about the complex acoustic world. Our system endows LALMs with universal spatial understanding through four key innovations: (1) A scalable simulation pipeline that synthesizes high-quality First-Order-Ambisonics(FOA) data; (2) A unified model framework that integrates universal spatial encoding with a dense hybrid projection mechanism to bridge the modality gap; (3) A progressive training curriculum that evolves from representation alignment to reinforcement learning-based reasoning; and (4) A comprehensive benchmark for audio scene analysis (ASA) designed to rigorously evaluate atomic perception, relational integration, and cognitive reasoning capabilities, on which our model demonstrates comparatively strong capability for spatial understanding. Our work provides a clear pathway for leveraging the powerful reasoning abilities of LALMs towards holistic ASA, advancing from "mono" semantic recognition to spatial intelligence.

Streaming TTS that receives streaming text is essential for interactive systems, yet this scheme faces two major challenges: unnatural prosody due to missing lookahead and long-form collapse due to unbounded context. We propose a prosodic-boundary-aware post-training strategy, adapting a pretrained LLM-based TTS model using weakly time-aligned data. Specifically, the model is adapted to learn early stopping at specified content boundaries when provided with limited future text. During inference, a sliding-window prompt carries forward previous text and speech tokens, ensuring bounded context and seamless concatenation. Evaluations show our method outperforms CosyVoice-Style interleaved baseline in both short and long-form scenarios. In long-text synthesis, especially, it achieves a 66.2% absolute reduction in word error rate (from 71.0% to 4.8%) and increases speaker and emotion similarity by 16.1% and 1.5% relatively, offering a robust solution for streaming TTS with incremental text.

Despite recent advances, synthetic voices often lack expressiveness due to limited prosody control in commercial text-to-speech (TTS) systems. We introduce the first end-to-end pipeline that inserts Speech Synthesis Markup Language (SSML) tags into French text to control pitch, speaking rate, volume, and pause duration. We employ a cascaded architecture with two QLoRA-fine-tuned Qwen 2.5-7B models: one predicts phrase-break positions and the other performs regression on prosodic targets, generating commercial TTS-compatible SSML markup. Evaluated on a 14-hour French podcast corpus, our method achieves 99.2% F1 for break placement and reduces mean absolute error on pitch, rate, and volume by 25-40% compared with prompting-only large language models (LLMs) and a BiLSTM baseline. In perceptual evaluation involving 18 participants across over 9 hours of synthesized audio, SSML-enhanced speech generated by our pipeline significantly improves naturalness, with the mean opinion score increasing from 3.20 to 3.87 (p < 0.005). Additionally, 15 of 18 listeners preferred our enhanced synthesis. These results demonstrate substantial progress in bridging the expressiveness gap between synthetic and natural French speech. Our code is publicly available at https://github.com/hi-paris/Prosody-Control-French-TTS.

Laughing, sighing, stuttering, and other forms of paralanguage do not contribute any direct lexical meaning to speech, but they provide crucial propositional context that aids semantic and pragmatic processes such as irony. It is thus important for artificial social agents to both understand and be able to generate speech with semantically-important paralanguage. Most speech datasets do not include transcribed non-lexical speech sounds and disfluencies, while those that do are typically multi-speaker datasets where each speaker provides relatively little audio. This makes it challenging to train conversational Text-to-Speech (TTS) synthesis models that include such paralinguistic components. We thus present DisfluencySpeech, a studio-quality labeled English speech dataset with paralanguage. A single speaker recreates nearly 10 hours of expressive utterances from the Switchboard-1 Telephone Speech Corpus (Switchboard), simulating realistic informal conversations. To aid the development of a TTS model that is able to predictively synthesise paralanguage from text without such components, we provide three different transcripts at different levels of information removal (removal of non-speech events, removal of non-sentence elements, and removal of false starts), as well as benchmark TTS models trained on each of these levels.

Despite the strong performance of large audio language models (LALMs) in various tasks, exactly how and where they integrate acoustic features with textual context remains unclear. We adapt causal tracing to investigate the internal information flow of LALMs during audio comprehension. By conducting layer-wise and token-wise analyses across DeSTA, Qwen, and Voxtral, we evaluate the causal effects of individual hidden states. Layer-wise analysis identifies different fusion strategies, from progressive integration in DeSTA to abrupt late-stage fusion in Qwen. Token-wise analysis shows that the final sequence token acts as an informational bottleneck where the network decisively retrieves relevant information from the audio. We also observe an attention-like query mechanism at intermediate token positions that triggers the model to pull task-relevant audio context. These findings provide a clear characterization of when and where multi-modal integration occurs within LALMs.

Multimodal large language models can exhibit text dominance, over-relying on linguistic priors instead of grounding predictions in non-text inputs. One example is large audio-language models (LALMs) where decisive audio evidence can be under-utilized even when it contains important information. To address this issue we use mechanistic interpretability to identify a small set of audio-specialist attention heads whose audio attention yields a ``listening'' signal. We show that this signal increases when audio evidence affects the model's output, providing an indicator of audio engagement under standard prompting. Leveraging this localization, we construct an audio--silence steering direction and apply an inference-time activation intervention to the final representation, amplifying the model's audio effect. To demonstrate the utility of this intervention, we show on MMAU that this improves accuracy by up to +8.0 percentage points on two Qwen-based LALMs, without any parameter updates.

The cloning of a speaker's voice using an untranscribed reference sample is one of the great advances of modern neural text-to-speech (TTS) methods. Approaches for mimicking the prosody of a transcribed reference audio have also been proposed recently. In this work, we bring these two tasks together for the first time through utterance level normalization in conjunction with an utterance level speaker embedding. We further introduce a lightweight aligner for extracting fine-grained prosodic features, that can be finetuned on individual samples within seconds. We show that it is possible to clone the voice of a speaker as well as the prosody of a spoken reference independently without any degradation in quality and high similarity to both original voice and prosody, as our objective evaluation and human study show. All of our code and trained models are available, alongside static and interactive demos.

It is known that human speech and certain animal vocalizations can convey meaningful content because we can decipher the content that a given utterance does convey. This paper explores an alternative approach to determining whether a signal is meaningful, one that analyzes only the signal itself and is independent of what the conveyed meaning might be. We devise a method that takes a waveform as input and outputs a score indicating its degree of `meaningfulness`. We cluster contiguous portions of the input to minimize the total description length, and then take the length of the code of the assigned cluster labels as meaningfulness score. We evaluate our method empirically, against several baselines, and show that it is the only one to give a high score to human speech in various languages and with various speakers, a moderate score to animal vocalizations from birds and orcas, and a low score to ambient noise from various sources.

Despite significant advancements in neural text-to-audio generation, challenges persist in controllability and evaluation. This paper addresses these issues through the Sound Scene Synthesis challenge held as part of the Detection and Classification of Acoustic Scenes and Events 2024. We present an evaluation protocol combining objective metric, namely Fr\'echet Audio Distance, with perceptual assessments, utilizing a structured prompt format to enable diverse captions and effective evaluation. Our analysis reveals varying performance across sound categories and model architectures, with larger models generally excelling but innovative lightweight approaches also showing promise. The strong correlation between objective metrics and human ratings validates our evaluation approach. We discuss outcomes in terms of audio quality, controllability, and architectural considerations for text-to-audio synthesizers, providing direction for future research.

The automatic movie dubbing model generates vivid speech from given scripts, replicating a speaker's timbre from a brief timbre prompt while ensuring lip-sync with the silent video. Existing approaches simulate a simplified workflow where actors dub directly without preparation, overlooking the critical director-actor interaction. In contrast, authentic workflows involve a dynamic collaboration: directors actively engage with actors, guiding them to internalize the context cues, specifically emotion, before performance. To address this issue, we propose a new Retrieve-Augmented Director-Actor Interaction Learning scheme to achieve authentic movie dubbing, termed Authentic-Dubber, which contains three novel mechanisms: (1) We construct a multimodal Reference Footage library to simulate the learning footage provided by directors. Note that we integrate Large Language Models (LLMs) to achieve deep comprehension of emotional representations across multimodal signals. (2) To emulate how actors efficiently and comprehensively internalize director-provided footage during dubbing, we propose an Emotion-Similarity-based Retrieval-Augmentation strategy. This strategy retrieves the most relevant multimodal information that aligns with the target silent video. (3) We develop a Progressive Graph-based speech generation approach that incrementally incorporates the retrieved multimodal emotional knowledge, thereby simulating the actor's final dubbing process. The above mechanisms enable the Authentic-Dubber to faithfully replicate the authentic dubbing workflow, achieving comprehensive improvements in emotional expressiveness. Both subjective and objective evaluations on the V2C Animation benchmark dataset validate the effectiveness. The code and demos are available at https://github.com/AI-S2-Lab/Authentic-Dubber.

To support machine learning of cross-language prosodic mappings and other ways to improve speech-to-speech translation, we present a protocol for collecting closely matched pairs of utterances across languages, a description of the resulting data collection and its public release, and some observations and musings. This report is intended for: people using this corpus, people extending this corpus, and people designing similar collections of bilingual dialog data.

Developing high-quality text-to-speech (TTS) systems for low-resource languages is challenging due to the scarcity of paired text and speech data. In contrast, automatic speech recognition (ASR) models for such languages are often more accessible, owing to large-scale multilingual pre-training efforts. We propose a framework based on Group Relative Policy Optimization (GRPO) to adapt an autoregressive, multilingual TTS model to new languages. Our method first establishes a language-agnostic foundation for TTS synthesis by training a multilingual baseline with International Phonetic Alphabet (IPA) tokens. Next, we fine-tune this model on limited paired data of the new languages to capture the target language's prosodic features. Finally, we apply GRPO to optimize the model using only unpaired text and speaker prompts, guided by a multi-objective reward from pretrained ASR, speaker verification, and audio quality estimation models. Experiments demonstrate that this pipeline produces intelligible and speaker-consistent speech in low-resource languages, substantially outperforming fine-tuning alone. Furthermore, our GRPO-based framework also improves TTS performance in high-resource languages, surpassing offline alignment methods such as Direct Preference Optimization (DPO) yielding superior intelligibility, speaker similarity, and audio quality.

We propose a novel text-to-speech (TTS) framework centered around a neural transducer. Our approach divides the whole TTS pipeline into semantic-level sequence-to-sequence (seq2seq) modeling and fine-grained acoustic modeling stages, utilizing discrete semantic tokens obtained from wav2vec2.0 embeddings. For a robust and efficient alignment modeling, we employ a neural transducer named token transducer for the semantic token prediction, benefiting from its hard monotonic alignment constraints. Subsequently, a non-autoregressive (NAR) speech generator efficiently synthesizes waveforms from these semantic tokens. Additionally, a reference speech controls temporal dynamics and acoustic conditions at each stage. This decoupled framework reduces the training complexity of TTS while allowing each stage to focus on semantic and acoustic modeling. Our experimental results on zero-shot adaptive TTS demonstrate that our model surpasses the baseline in terms of speech quality and speaker similarity, both objectively and subjectively. We also delve into the inference speed and prosody control capabilities of our approach, highlighting the potential of neural transducers in TTS frameworks.

Large language models (LLMs) have recently advanced auditory speech recognition (ASR), visual speech recognition (VSR), and audio-visual speech recognition (AVSR). However, understanding of their internal dynamics under fine-tuning remains limited. In natural language processing, recent work has revealed attention sinks, tokens that attract disproportionately high attention, and associated massive activations in which some features of sink tokens exhibit huge activation in LLMs. In this work, we are the first to study these phenomena in multimodal speech recognition. Through a detailed analysis of audio-visual LLMs, we identify attention sinks and massive activations not only at the BOS token but also at intermediate low-semantic tokens across ASR, VSR, and AVSR. We show that massive activations originate in the MLP layers and correspond to fixed feature indices across all sink tokens. We further show that intermediate sink tokens exhibit high cosine similarity to the BOS token, thereby amplifying attention and activation. Building on these insights, we introduce a simple decorrelation loss that reduces cosine similarity between BOS and other tokens, effectively mitigating intermediate sinks and massive activations. Furthermore, our method improves word error rate (WER) under high audio-visual feature downsampling while remaining stable at lower downsampling rates.