Trelis/latent-space-train
收藏Hugging Face2026-01-31 更新2026-02-07 收录
下载链接：
https://hf-mirror.com/datasets/Trelis/latent-space-train
下载链接
链接失效反馈官方服务：
资源简介：
这是一个用于Whisper模型训练的语音数据集，由Trelis Studio准备。数据集包含音频片段及其文本转录，特别提供了带时间戳的转录文本。音频片段经过VAD处理，去除了静音部分，仅保留语音区域。数据集还包含音频片段的开始和结束时间、语音持续时间、词级时间戳等信息。训练建议采用两桶方法，即50%使用不带时间戳的纯文本转录，50%使用带Whisper时间戳标记的转录文本。
Speech dataset prepared with Trelis Studio for Whisper model training. The dataset includes audio segments and their text transcriptions, with special provision for transcriptions with timestamp tokens. Audio segments are processed with VAD to remove silence, retaining only speech regions. The dataset also includes segment start and end times, speech duration, word-level timestamps, etc. Training usage recommends a two-bucket approach: 50% with plain transcriptions (no timestamps) and 50% with transcriptions containing Whisper timestamp tokens.
提供机构：
Trelis
