Overview
LibriSpeech is the foundational benchmark dataset for automatic speech recognition, derived from LibriVox audiobook recordings. Its 1,000 hours of clean English read speech with aligned transcripts have driven ASR progress for a decade.
What’s In It
LibriSpeech contains 1,000 hours of read English speech sampled at 16kHz from LibriVox audiobooks. The data is split into clean and other (noisier) subsets, each divided into train, dev, and test portions with full text transcriptions.
How It’s Used
LibriSpeech word error rate is the standard metric for ASR evaluation. Whisper, Wav2Vec 2.0, and virtually every modern ASR system reports scores on it. Pre-training approaches use the audio for self-supervised speech representation learning.
Controversies
As audiobook data, LibriSpeech represents only read speech with limited speaker diversity. The predominantly American English content limits multilingual evaluation. Modern models have largely saturated the benchmark with WER below 2%.