LibriSpeech is a corpus of approximately 1,000 hours of read English speech derived from public-domain audiobooks in the LibriVox project, with transcripts drawn from the corresponding Project Gutenberg texts. Audio is segmented and force-aligned to the text and provided at 16 kHz, making it a clean, ready-to-use resource for training and evaluating speech models, and it is accompanied by pronunciation lexicons and language-model training material.
What’s in it
The data is partitioned into standard splits: train-clean-100, train-clean-360, and train-other-500 for training, plus matched dev-clean/dev-other and test-clean/test-other evaluation sets. The clean versus other division reflects recording quality and speaker-accent difficulty, letting researchers report results on easier and harder conditions separately, and speakers are kept disjoint across splits to prevent leakage. A larger multilingual sibling, Multilingual LibriSpeech (MLS), and the unlabeled Libri-Light set extend the same source material for cross-lingual and self-supervised work.
Tasks, license & access
LibriSpeech is one of the most widely used benchmarks for automatic speech recognition (ASR), commonly reported via word error rate, and it also serves acoustic modeling, self-supervised speech pretraining, and speaker research. It was prepared by researchers associated with Johns Hopkins University (Panayotov, Chen, Povey, and Khudanpur) and is released under a permissive CC BY 4.0 license, freely downloadable from OpenSLR.
- Modality: read English speech audio with aligned transcripts
- Scale: ~1,000 hours, 16 kHz
- Tasks: ASR, acoustic modeling, self-supervised speech pretraining
- License: CC BY 4.0
- Creator: Johns Hopkins University researchers (2015)
Its size, clean alignment, and open license made LibriSpeech a default starting point for English ASR and a common testbed for modern speech representation learning.