Skip to content
RAG Repo

LibriSpeech

LibriSpeech is the reference dataset for English automatic speech recognition (ASR, turning spoken audio into text). It draws on LibriVox, a collection of volunteer-read public domain audiobooks, and pairs the audio with sentence-level transcripts so models can learn the mapping from sound to words.

The corpus is split into training, development, and test sets, with clean and noisier subsets so you can measure how a model holds up in harder conditions. Audio is stored as FLAC (a lossless open format), and the whole set totals around 1,000 hours. It remains the most common yardstick for reporting English ASR accuracy in research papers.

speechenglishasrpublic-domainbenchmarkaudiobooks

Related sources