HiFiTTS-2
收藏资源简介:
HiFiTTS-2是一个大规模的语音数据集,用于高带宽语音合成。该数据集源于LibriVox有声读物,包含约36.7k小时的英语语音用于22.05 kHz的训练,以及31.7k小时用于44.1 kHz的训练。数据集还包括详细的语句和有声读物元数据,以便研究人员根据不同的用例对数据集进行质量筛选。实验结果表明,我们的数据管道和生成的数据集可以促进高质量、零样本文本到语音(TTS)模型在高带宽下的训练。该数据集适用于语音合成和语音克隆等应用,并为研究高带宽和混合带宽建模提供了新的数据资源。
HiFiTTS-2 is a large-scale speech dataset dedicated to high-bandwidth speech synthesis. Derived from LibriVox audiobooks, this dataset contains approximately 36.7k hours of English speech for training at 22.05 kHz, and 31.7k hours for training at 44.1 kHz. It also includes detailed utterance and audiobook metadata, enabling researchers to perform quality filtering on the dataset according to different use cases. Experimental results demonstrate that our data pipeline and the generated dataset can facilitate the training of high-quality zero-shot text-to-speech (TTS) models for high-bandwidth scenarios. This dataset is suitable for applications such as speech synthesis and voice cloning, and provides a new data resource for research on high-bandwidth and mixed-bandwidth modeling.
HiFiTTS-2 数据集概述
数据集基本信息
- 名称: HiFiTTS-2: A Large-Scale High Bandwidth Speech Dataset
- 所有者: NVIDIA Corporation
- 创建日期: June 2025
- 许可证: Creative Commons Attribution 4.0 International License (CC BY 4.0)
- 论文链接: https://arxiv.org/abs/2506.04152
数据集内容
- 来源: LibriVox 有声读物
- 音频时长: 约 36.7k 小时
- 说话人数量: 5k
- 采样率: 48 kHz(原始下载)
- 配置:
- 22kHz: 适用于 22 kHz 训练的基数据集
- 44kHz: 预计算子集,31.7k 小时,适用于 44 kHz 训练
数据集格式
- 字段:
- utterance level manifest:
- audio_filepath, speaker, set, duration, bandwidth, speaker_count, wer, cer, text_source, text, normalized_text
- audiobook chapter level manifest:
- url, chapter_filepath, bandwidth, utterances (包含 audio_filepath, offset, duration)
- utterance level manifest:
下载说明
- 下载元数据文件: 从 Hugging Face 仓库下载 manifest.json 和 chapter.json 文件。
- 安装工具: 安装 NeMo-speech-data-processor (SDP)。
- 运行下载脚本: bash python /home/NeMo-speech-data-processor/main.py --config-path="/home/NeMo-speech-data-processor/dataset_configs/english/hifitts2" --config-name="config_22khz.yaml" workspace_dir="/home/hifitts2" max_workers=8
- 磁盘空间需求:
- 22 kHz 版本: 约 3TB
- 44 kHz 版本: 约 5TB
引用
bibtex @inproceedings{rlangman2025hifitts2, title={HiFiTTS-2: A Large-Scale High Bandwidth Speech Dataset}, author={Ryan Langman and Xuesong Yang and Paarth Neekhara and Shehzeen Hussain and Edresson Casanova and Evelina Bakhturina and Jason Li}, booktitle={Interspeech}, year={2025}, }




