WordVoice-5A Dataset 🚀

A Large-Scale Bilingual Word-level Five-Annotation Dataset for WordVoice

---

📖 Dataset Description / 数据集简介

WordVoice-5A is a large-scale bilingual (Mandarin and English) dataset containing approximately 4.7k hours of speech with fine-grained word-level acoustic annotations, designed for high-precision controllable Text-to-Speech (TTS). It addresses the scarcity of large-scale, high-quality word-aligned datasets with explicit acoustic annotations in the open-source community. Through a linguistically guided automatic annotation pipeline based on empirical acoustic distributions, the dataset provides five core acoustic attributes for each word: Duration, Acoustic Boundary, Energy, Pitch, and Tone. It aims to break the "black-box" nature of LLM-based TTS and facilitate future research in fine-grained speech generation and acoustic modeling.

WordVoice-5A 是一个约 4.7k 小时的大规模中英双语字/词级声学属性标注数据集，专为高精度、细粒度可控语音合成（TTS）设计。该数据集针对当前开源社区缺乏大规模、高质量字/词级对齐与显式声学标注数据的问题，提出了一套结合语言学规则与真实声学统计分布的自动化标注 Pipeline，为每个字/词提供 时长（Duration）、声学边界（Acoustic Boundary）、能量（Energy）、音高（Pitch）和音调（Tone） 五维核心属性。WordVoice-5A 致力于打破 LLM-TTS 的"黑盒"特性，推动细粒度语音生成与声学建模研究。

---

🔗 WordVoice Ecosystem / WordVoice 生态系统

This dataset is a core component of the WordVoice project. We also provide the official data processing pipeline and the pre-trained TTS models: 本数据集是 WordVoice 项目的核心部分。我们同时开源了官方的数据处理流水线与预训练 TTS 模型：

• 🛠️ WordVoice Data Pipeline: The linguistically-guided automated annotation toolkit used to build this dataset. / 用于构建本数据集的语言学指导自动化标注工具包。
• 🧠 WordVoice Model: The official implementation of the WordVoice TTS framework, supporting explicit multi-dimensional word-level control. / WordVoice TTS 框架的官方实现，支持显式的多维字级控制。

---

📊 Dataset Summary / 数据集概述

The WordVoice-5A dataset contains approximately 4,684 hours of high-quality speech, including 2,546 hours of Mandarin and 2,138 hours of English, with over 52 million word-level annotations. The raw speech and text data are sourced from the open-source LEMAS corpus. Our primary contribution lies in the comprehensive data annotation rather than data cleansing. By applying our rigorous dual-model alignment and linguistically guided pipeline to the raw data, we successfully extracted and annotated five-dimensional acoustic attributes for every single word. It is specifically designed for:

• Controllable TTS: Training TTS models with explicit word-level acoustic control.
• Prosody Modeling: Studying bilingual micro-prosody and coarticulation in continuous speech.

WordVoice-5A 数据集包含约 4,684 小时高质量语音，其中中文 2,546 小时、英文 2,138 小时，共包含超过 5,200 万字/词级标注。本数据集的原始语音与文本数据来源于开源的 LEMAS 语料库。我们的核心工作是对原始数据进行了深度的数据标注（而非单纯的数据清洗）。通过严格的双模型交叉对齐与语言学指导的自动化 Pipeline，我们成功为原始语料中的每一个字/词提取并标注了五维声学属性。该数据集主要面向以下研究方向：

• 可控语音合成（Controllable TTS）：训练支持字/词级声学属性显式控制的 TTS 模型。
• 韵律建模（Prosody Modeling）：研究中英双语连续语流中的微观韵律及协同发音规律。

---

✨ Key Features / 主要特点

• 5-Dimensional Word-Level Annotations / 五维字级标注
  Each character/word is annotated with Duration, Boundary, Energy, Pitch, and Tone.
  每个字/词均包含精准的 时长（Duration）、5级声学边界（Acoustic Boundary）、能量（Energy）、音高（Pitch） 和 7类音调（Tone） 标注。

• Massive Bilingual Corpus / 超大规模双语语料
  The dataset contains 2,546 hours of Mandarin and 2,138 hours of English, making it one of the largest publicly available corpora with comprehensive word-level acoustic annotations.
  数据集包含 2546 小时中文 与 2138 小时英文，是目前已知规模最大、标注维度最完整的字/词级控制数据集之一。

• Linguistically Guided Annotation / 语言学专家指导标注
  Annotation criteria (e.g., coarticulation-aware truncation and quadratic pitch contour fitting) are carefully designed based on empirical acoustic distributions and linguistic principles.
  标注标准与阈值（如"掐头去尾"去除协同发音、二次曲线拟合音调等）均基于真实数据分布与语言学规则设计。

• High-Precision Timestamps / 高精度时间戳
  Word boundaries are obtained through dual-model alignment (MFA & Qwen3FA) and refined by loudness-based boundary optimization to ensure high alignment fidelity.
  采用 MFA 与 Qwen3FA 双模型交叉验证，并结合基于响度的边界优化策略，保证字/词级时间戳的高精度。

---

📂 Data Structure & Annotation Details / 数据结构与标注说明

The dataset is provided in JSONL format paired with corresponding audio files. Each record contains the audio path, transcript, and word-level acoustic annotations.

数据集采用 JSONL 格式，并配有对应音频文件。每条记录包含音频路径、文本以及对应字/词级五维声学属性标注。

Data Format Example / 数据格式示例

{
  "utt": "zh_WenetSpeech4TTS_0001681467",
  "audio_path": "test/WenetSpeech4TTS_0001681467.mp3",
  "duration": 1.92,
  "text": "真是巧啊。",
  "mfa_text": "真 是 巧 啊",
  "mfa_words": [
    {"word": "真", "start": 0.69, "end": 0.84},
    {"word": "是", "start": 0.84, "end": 0.95},
    {"word": "巧", "start": 0.95, "end": 1.16},
    {"word": "啊", "start": 1.17, "end": 1.34}],
  "f0": [-0.2686, -0.246, -0.3819, -0.6745],
  "eng": [0.5415, 0.4326, 0.3464, 0.2612],
  "tone": ["flat", "flat", "fall", "fall"],
  "bnd": ["b0", "b0", "b1", "b4"]
}

Annotation Ranges / 标注范围说明

• Duration（时长）

  • Float (seconds)
  • 字/词的实际发音时长（单位：秒）。

• Boundary（声学边界）

  • Five discrete categories representing the pause level after each word.
  • 表示字/词后的停顿等级，共 5 类：
    • b0: No pause / 无停顿
    • b1: ≤ 0.05 s (Micro pause / 微停顿)
    • b2: ≤ 0.18 s (Word boundary / 词边界)
    • b3: ≤ 0.40 s (Comma-level boundary / 逗号级边界)
    • b4: > 0.40 s (Sentence boundary / 句号级边界)

• Energy（能量）

  • Float normalized to [0, 1]
  • 字/词级归一化有效响度。

• Pitch（音高）

  • Float normalized to [-1, 1]
  • 字/词级核心音高均值。

• Tone（音调）

  • Seven discrete pitch contour categories.
  • 字内音高变化轮廓，共 7 类：
    • flat
    • rise
    • rrise
    • fall
    • ffall
    • peak
    • valley

---

🎯 Use Cases / 使用场景

• Fine-grained controllable LLM-based TTS

• Word-level prosody modeling

• Local prosody editing for audiobook narration and dubbing

• Cross-lingual acoustic feature analysis

• Prosody prediction

• 细粒度可控 LLM-TTS 模型训练

• 字/词级韵律建模

• 有声书与视频配音中的精准局部韵律编辑

• 跨语种字/词级声学特征分析

• 韵律预测

---

📝 Citation / 引用

If you find this dataset useful in your research, please cite our paper:

@misc{nie2026wordvoice,
      title={WordVoice: Explicit and Decoupled Multi-Dimensional Word-Level Control for LLM-Based TTS}, 
      author={Sihang Nie and Jinxin Ji and Xiaofen Xing and Deyi Tuo and Chengbin Jin and Jialong Mai and Xiangmin Xu},
      year={2026},
      eprint={2607.06461},
      archivePrefix={arXiv},
      primaryClass={eess.AS},
      url={https://arxiv.org/abs/2607.06461}, 
}