WordVoice-5A: A Large-Scale Bilingual Dataset for Fine-Grained Controllable TTS
Dataset Description / 数据集简介
WordVoice-5A 是一个约 4.7k 小时的大规模中英双语字级声学属性标注数据集，专为高精度、细粒度可控的语音合成（TTS）设计。该数据集解决了当前开源社区缺乏大规模、高质量字级对齐与声学标注数据的问题。通过融合语言学规则与声学特征统计分布的自动化标注 Pipeline，我们提取了五维核心属性（时长、边界、能量、音高、音调）。该数据集旨在打破 LLM-TTS 的“黑盒”特性，推动精细化语音生成与声学建模的未来研究。
WordVoice-5A is a large-scale, ~4.7k-hour bilingual (Mandarin and English) dataset with fine-grained word-level acoustic annotations, designed for high-precision, controllable Text-to-Speech (TTS). It addresses the critical scarcity of massive, high-quality word-aligned and acoustically annotated data in the open-source community. Processed through a linguistically-guided automated pipeline based on empirical acoustic distributions, it provides five-dimensional core attributes (Duration, Boundary, Energy, Pitch, and Tone). This dataset aims to break the "black-box" nature of LLM-TTS and facilitate future research in fine-grained speech generation and acoustic modeling.
Dataset Summary / 数据集概述
WordVoice-5A 数据集包含约 4,684 小时的高质量语音数据（中文 2,546 小时，英文 2,138 小时），涵盖超过 5,200 万个字/词级标注。数据源自开源 LEMAS 数据集，经过严格的双模型交叉验证与清洗。数据集专门设计用于：
- **可控语音合成 (Controllable TTS)**：训练支持字级声学属性显式控制的 TTS 模型。
- **韵律建模 (Prosody Modeling)**：研究中英双语在连续语流中的微观韵律与协同发音规律。
The WordVoice-5A dataset contains approximately 4,684 hours of high-quality speech data (2,546 hours of Mandarin, 2,138 hours of English), encompassing over 52 million character/word-level annotations. Derived from the open-source LEMAS dataset and refined through rigorous dual-model cross-validation and cleansing, it is specifically designed for:
- Controllable TTS: Training TTS models that support explicit word-level acoustic control.
- Prosody Modeling: Researching micro-prosody and coarticulation patterns in continuous bilingual speech.
Key Features / 主要特点
- **五维字级标注 (5-Dimensional Annotations)**：每个字/词均包含精准的时长 (Duration)、5级声学边界 (Boundary)、能量 (Energy)、音高 (Pitch) 与 7类音调 (Tone) 标注。
- **超大规模双语语料 (Massive Bilingual Corpus)**：包含中文 2546h 与英文 2138h，是目前已知规模最大、标注维度最全的字级控制数据集。
- **语言学专家指导 (Linguistically-Guided)**：分类标准与阈值（如“掐头去尾”去除协同发音、二次曲线拟合音调）均基于真实数据分布与语言学规则严格论证。
- **高精度时间戳 (High-Precision Timestamps)**：采用 MFA 与 Qwen3FA 双模型交叉验证，结合基于响度的边界优化，严格剔除不良对齐数据。
English:
- 5-Dimensional Word-Level Annotations: Every character/word is annotated with precise Duration, 5-level Acoustic Boundary, Energy, Pitch, and 7-category Tone.
- Massive Bilingual Corpus: Comprising 2,546h of Mandarin and 2,138h of English, making it the largest and most comprehensively annotated dataset for word-level control.
- Linguistically-Guided: Classification standards and thresholds (e.g., truncation for coarticulation mitigation, quadratic curve fitting for Tone) are strictly driven by empirical data distributions and linguistic rules.
- High-Precision Timestamps: Utilizes dual-model (MFA & Qwen3FA) cross-validation combined with loudness-based boundary optimization to ensure alignment fidelity.
Use Cases / 使用场景
- 细粒度可控 LLM-TTS 模型训练 (Training fine-grained controllable LLM-TTS models)
- 有声书演播与视频配音中的精准局部韵律编辑 (Precise local prosody editing for audiobook narration and video dubbing)
- 跨语种字级声学特征分析与韵律预测 (Cross-lingual word-level acoustic feature analysis and prosody prediction)
Citation / 引用
If you find this dataset useful in your research, please cite our paper:
@inproceedings{wordvoice2027,
title={WordVoice: Explicit and Decoupled Multi-Dimensional Word-Level Control for LLM-Based TTS},
author={Your Name and Co-authors},
booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
year={2027}
}