LibriQuote
收藏资源简介:
LibriQuote 是一个由 Deezer Research 和 LORIA-CNRS 共同创建的英语语料库,从有声读物中提取,旨在用于细调和评估具有表现力的零样本语音合成系统。该数据集包含 12.7K 小时的非表现性语音和 5.3K 小时的主要表现性语音,每个表现性语音都附有上下文信息和描述语音的动词和副词的伪标签。此外,还提供了一个具有挑战性的 7.5 小时测试集,用于评估语音合成系统的表现力。LibriQuote 数据集对于训练和评估表现力语音合成系统具有重要意义,可以促进语音合成技术的进一步发展。
LibriQuote is an English corpus co-created by Deezer Research and LORIA-CNRS, extracted from audiobooks, and designed for fine-tuning and evaluating expressive zero-shot speech synthesis systems. The dataset contains 12.7K hours of non-expressive speech and 5.3K hours of predominantly expressive speech, where each expressive speech utterance is paired with contextual information and pseudo-labels of verbs and adverbs describing the speech. In addition, a challenging 7.5-hour test set is provided for evaluating the expressiveness of speech synthesis systems. The LibriQuote dataset is of great significance for training and evaluating expressive speech synthesis systems, and can promote the further development of speech synthesis technology.
LibriQuote 数据集概述
数据集简介
LibriQuote 是一个用于表达性零样本语音合成的虚构角色语音数据集,包含从小说中提取的角色对话和叙述段落语音片段。
核心内容
- 数据来源:基于 LibriVox 录音,使用 LibriLight 音频文件作为后端音频文件
- 音频格式:16KHz 采样率
- 数据组成:包含叙述段落和小说角色引语的语音片段
数据集获取
- 主存储位置:https://huggingface.co/datasets/gasmichel/LibriQuote
- 测试音频样本:https://huggingface.co/datasets/gasmichel/LibriQuote/tree/main/test_audios
相关资源
- 论文地址:https://arxiv.org/pdf/2509.04072
- 音频样本演示:https://libriquote.github.io/
- 原始音频文件下载指引:https://github.com/facebookresearch/libri-light/blob/main/data_preparation/README.md
数据处理与评估
- 处理工具:提供 Python 工具类处理 LibriQuote 数据
- 评估脚本:包含用于在 LibriQuote-test 上基准测试 TTS 系统的评估脚本
引用信息
bibtex @misc{Michel2025LibriQuote, title={LibriQuote: A Speech Dataset of Fictional Character Utterances for Expressive Zero-Shot Speech Synthesis}, author={Gaspard Michel and Elena V. Epure and Christophe Cerisara}, year={2025}, eprint={2509.04072}, archivePrefix={arXiv}, primaryClass={eess.AS}, url={https://arxiv.org/abs/2509.04072} }




