WenetSpeech

WenetSpeech 是一个多领域的普通话语料库，由 10,000+ 小时的高质量标记语音、2,400+ 小时的弱标记语音和约 10,000 小时的未标记语音组成，总共 22,400+ 小时。作者从 YouTube 和 Podcast 收集了数据，这些数据涵盖了各种说话风格、场景、领域、主题和嘈杂的条件。引入了一种基于光学字符识别 (OCR) 的方法来为 YouTube 数据在其相应的视频字幕上生成音频/文本分割候选。

WenetSpeech is a multi-domain Mandarin speech corpus. It consists of over 10,000 hours of high-quality manually labeled speech, over 2,400 hours of weakly labeled speech, and approximately 10,000 hours of unlabeled speech, with a total duration of over 22,400 hours. The authors collected the dataset from YouTube and Podcast, which covers diverse speaking styles, scenarios, domains, topics, and noisy conditions. An optical character recognition (OCR)-based method was introduced to generate audio-text segmentation candidates for the YouTube data using its corresponding video subtitles.

搜集汇总

背景与挑战

背景概述

WenetSpeech是一个包含22,400+小时的多领域普通话语料库，涵盖高质量标记、弱标记和未标记语音数据，数据来源于YouTube和Podcast，适用于语音识别研究。