0x3/librispeech_asr

LibriSpeech是一个包含约1000小时、16kHz采样率的英语朗读语音语料库，由Vassil Panayotov在Daniel Povey的协助下创建。数据来源于LibriVox项目的朗读有声读物，并经过仔细的分段和对齐处理。数据集主要用于自动语音识别(ASR)和说话人识别任务，支持两种配置：clean（包含录音质量较高、口音更接近美式英语的语音，分为train.100（100小时，28,539个样本）、train.360（360小时，104,014个样本）、validation（2,703个样本）和test（2,620个样本））和other（包含其他语音，分为train.500（500小时，148,688个样本）、validation（2,864个样本）和test（2,939个样本）），以及一个all配置合并所有数据。每个数据点包括音频文件路径、音频数据（以16kHz采样率解码）、转录文本、说话人ID、章节ID和唯一ID。数据集采用CC BY 4.0许可证，基于公共领域音频书籍构建，常用于评估模型在单词错误率(WER)等指标上的性能。

LibriSpeech is a corpus of approximately 1000 hours of 16kHz read English speech, prepared by Vassil Panayotov with the assistance of Daniel Povey. The data is derived from read audiobooks from the LibriVox project, and has been carefully segmented and aligned. It is designed for automatic speech recognition (ASR) and speaker identification tasks, with configurations including clean (higher recording quality and accents closer to US English, with splits: train.100 (100 hours, 28,539 examples), train.360 (360 hours, 104,014 examples), validation (2,703 examples), and test (2,620 examples)) and other (other speech, with splits: train.500 (500 hours, 148,688 examples), validation (2,864 examples), and test (2,939 examples)), plus an all config combining all data. Each data instance includes the audio file path, audio data (decoded at 16kHz sampling rate), transcription text, speaker ID, chapter ID, and unique ID. The dataset is licensed under CC BY 4.0, built from public domain audio books, and commonly used for evaluating model performance on metrics like word error rate (WER).