HKUSTAudio/ISCSLP2026-CoT-TTS

ISCSLP 2026 CoT-TTS Dataset

Dataset Overview

This dataset is prepared for the ISCSLP 2026 CoT-TTS Challenge and is designed to support research on context-aware, expressive, and CoT-guided speech generation. It is constructed from speech-rich media sources, including films, TV dramas, radio dramas, and short dramas, where dialogue often contains rich conversational context, speaker interactions, scene changes, and emotional variation. Each sample is organized around a target utterance, together with its preceding dialogue context, reference speech, metadata, and corresponding annotations, so that models can learn to infer an appropriate speaking style from context rather than relying only on the target text.

Language	Duration	Ratio	Segments
English	~8.6K h	54%	~1.62M
Chinese	~7.4K h	46%	~1.38M
Total	~16K h	100%	~3.0M

The released audio files are extracted segments from the original source recordings. To reduce storage usage, only the audio file format has been standardized to FLAC, while other acoustic properties are preserved as much as possible. The audio has not been aggressively standardized or denoised. As a result, different files may have different sampling rates, channel configurations, loudness levels, background sounds, or environmental noise. This design is intentional: the dataset aims to retain realistic acoustic conditions and avoid excessive preprocessing assumptions, allowing users and participants to decide how to process, normalize, enhance, or filter the audio according to their own methods. However, the metadata and annotations provided in this dataset were generated based on normalized and denoised versions of the corresponding audio segments, in order to improve annotation reliability and accuracy.

File Structure

HKUSTAudio/ISCSLP2026-CoT-TTS/
├── movie-en1/
│   ├── metadata.json
│   ├── continuous_segments/
│   │   └── *.flac
│   └── dialogue_segments/
│       └── *.flac
└── movie-zh2/
    ├── metadata.json
    ├── continuous_segments/
    │   └── *.flac
    └── dialogue_segments/
        └── *.flac

The dataset is divided into six folders. These folders follow the same internal structure and mainly differ in language and data partition. The split is used for easier upload, storage, and processing.

Each folder contains one metadata JSON file and two audio directories: dialogue_segments/ and continuous_segments/.

• dialogue_segments/ contains all utterance-level audio segments required by the metadata, including target speech, reference speech, and dialogue-context speech. These segments are extracted according to the utterance timestamps obtained after speaker diarization.

• continuous_segments/ also contains all required audio segments, including target speech, reference speech, and dialogue-context speech. Compared with dialogue_segments/, these segments are extended before and after each utterance to better preserve the continuity of the original audio context.

For example, if an original source span covers 0-20 seconds, dialogue_segments/ may contain only the detected speech intervals, such as 0-3s, 6-7s, and 13-15s. In contrast, continuous_segments/ may contain extended and continuous intervals, such as 0-4.5s, 4.5-9.5s, and 9.5-20s. This design allows users to choose between cleaner utterance-level segments and more context-preserving continuous audio segments.

Data Format

{
  "movie_id": "ID of the source media item",
  "movie_name": "Name of the source media file",
  "scene_index": "Index of the scene",
  "start_time": "Start time of the dialogue context",
  "end_time": "End time of the dialogue context",
  "dialog_count": "Number of historical dialogue segments",
  "target_segment": {
    "segment_id": "Unique ID of the target utterance",
    "ref_segment_id": "Segment ID of the reference speech",
    "text": "Text content of the target utterance",
    "speaker": "Original speaker label",
    "normalized_speaker": "Normalized speaker label",
    "emotion_tag": "Emotional description of audio clips",

    "features": {
      "duration": "Total duration of the target audio segment",
      "active_duration": "Valid duration of audio clip (excluding silence)",
      "loudness": "Loudness of the target audio segment",
      "expressive_intensity": "Expressive intensity score of the target utterance"
    },
    "cot_text": "Chain-of-thought style analysis",
    "start": "Start time of the target utterance",
    "end": "End time of the target utterance",
    "best_ref_similarity": "Similarity between the target utterance and the reference segment"
  },
  "dialog_segments": [
    {
      "segment_id": "Unique ID of a historical dialogue segment",
      "text": "Text content of the historical utterance",
      "speaker": "Original speaker label",
      "normalized_speaker": "Normalized speaker label",
      "emotion_tag": "Emotional description of audio clips",
      "features": {
      "duration": "Total duration of the target audio segment",
      "active_duration": "Valid duration of audio clip (excluding silence)",
      "loudness": "Loudness of the target audio segment",
      "expressive_intensity": "Expressive intensity score of the target utterance"
     },

      "start": "Start time of the historical utterance",
      "end": "End time of the historical utterance"
    },
      ...
  ],
  "index": "Unique sample index used for data loading and file matching"
}

Notes:

• All time values are in seconds.
• speaker: Original diarization label assigned within the full source episode.
• normalized_speaker: Speaker label re-indexed within the current scene or dialogue context.
• Audio files can be matched using segment_id, ref_segment_id.
• dialog_segments are ordered according to their temporal order in the original source media.

How to Use

Each sample is described by a JSON metadata entry. Users can read the JSON file to obtain the target utterance, historical dialogue context, speaker labels, emotion tags, CoT-style analysis, acoustic features, and timing information.

The corresponding audio files can be located using the segment IDs provided in the metadata. Specifically, segment_id identifies each utterance-level audio segment, while ref_segment_id identifies the reference speech segment used for speaker timbre.

A typical usage process is:

1. Load the JSON metadata.
2. Read dialog_segments as the historical dialogue context.
3. Read target_segment.text as the target text.
4. Use target_segment.ref_segment_id to locate the reference audio.
5. Use target_segment.segment_id to locate the target audio.
6. Use target_segment.cot_text as the reasoning-style annotation if needed.

Out-of-Scope Use

This dataset is released only for non-commercial academic research, development, and evaluation related to context-aware speech generation and CoT-guided TTS. The following uses are not permitted:

• Commercial use of the dataset, including commercial model training, commercial speech generation services, or commercial voice-cloning products.
• Redistribution, re-hosting, sublicensing, selling, or repackaging of the dataset or any part of it.
• Reconstruction, restoration, or redistribution of the original source media, including original films, dramas, audio programs, or other copyrighted materials.
• Attempts to identify, trace, or disclose the original speakers, actors, characters, source titles, or copyright holders from the released segments.
• Use of the dataset for impersonation, identity spoofing, deceptive speech generation, misinformation, fraud, harassment, or other harmful applications.
• Use of the dataset in systems intended for surveillance, speaker identification, biometric profiling, or privacy-invasive analysis.
• Any use that violates applicable laws, copyright regulations, platform policies, or the data usage terms provided by the dataset maintainers.

Limitations

This dataset is constructed from speech-rich media sources and therefore has several limitations that users should be aware of:

• The dataset may reflect the style, distribution, and dramatic characteristics of media content, and may not fully represent natural daily conversations.
• The released audio segments preserve many acoustic properties of the original recordings, so background sounds, music, environmental noise, overlapping speech, channel differences, loudness variation, and sampling-rate differences may exist.
• Speaker labels are automatically produced and should be treated as local anonymous labels, not as real speaker identities.
• Transcriptions, speaker diarization results, timestamps, emotion tags, and CoT-style annotations may contain errors.
• Emotion tags and reasoning annotations are intended as reference annotations and should not be treated as absolute ground truth.
• Some samples may contain culturally specific expressions, dramatic dialogue, or context-dependent speech styles.
• The dataset does not grant users any rights to the original copyrighted media beyond the limited research use of the processed dataset.
• Users are responsible for ensuring that their use of the dataset complies with applicable laws, institutional policies, and the dataset usage terms.

License and Copyright

This dataset is prepared for non-commercial academic research and evaluation purposes only. The source materials are collected from publicly accessible media sources. The dataset maintainers do not claim ownership of the original media content. The copyright of the original videos, films, dramas, audio programs, or other source materials remains with their respective copyright holders.

The released dataset does not include complete source media, full-length videos, full-length audio programs, or original media files. Instead, the released data consists of processed speech segments and corresponding metadata or annotations generated through a data preparation pipeline, including segmentation, clipping, normalization, filtering, transcription, speaker labeling, emotion annotation, and CoT-style annotation.

Users may use the dataset only for non-commercial research, development, and evaluation related to the intended research scope. Commercial use, redistribution, resale, sublicensing, public re-hosting, or any attempt to reconstruct, identify, or redistribute the original source materials is strictly prohibited.

The dataset is released under a non-commercial research-use license, following the spirit of the Creative Commons Attribution-NonCommercial 4.0 license. Users must follow the data usage terms provided by the dataset maintainers. The maintainers reserve the right to remove, modify, or restrict access to any data item if copyright, privacy, licensing, or ethical concerns are raised.

Access to the dataset does not transfer copyright ownership, neighboring rights, performer rights, publicity rights, or any other rights associated with the original source materials. Users are solely responsible for ensuring that their use of the dataset is lawful and appropriate in their jurisdiction.

Ethical Considerations

Users should handle this dataset responsibly and respect the rights of original content creators, performers, speakers, and copyright holders.

• Do not use the dataset to imitate, impersonate, or misrepresent any real person, actor, speaker, or character.
• Do not use the dataset to generate deceptive, harmful, defamatory, or misleading audio content.
• Do not attempt to identify speakers, recover source identities, or link released segments back to specific individuals or copyrighted works.
• Do not use the dataset to create systems that enable unauthorized voice cloning, biometric identification, surveillance, or privacy-invasive profiling.
• Clearly disclose the use of generated or synthetic speech when presenting outputs produced by models trained on this dataset.
• Respect all copyright, privacy, and licensing concerns related to the dataset and its source materials.
• If any data item raises copyright, privacy, or ethical concerns, please contact the dataset maintainers for review or takedown.

This dataset is provided to support research on context-aware and reasoning-guided speech generation. It should be used in a manner that is lawful, ethical, non-commercial, and respectful of the original rights holders.

Citation

TODO

Contact

For questions, issues, or requests related to the dataset, please contact:

• Weizhen Bian: wbian@connect.ust.hk