Libri-Light: A Benchmark for ASR with Limited or No Supervision

TL;DR

This paper presents Libri-Light, a large-scale, freely-available speech corpus derived from LibriVox audiobooks, and establishes benchmarks for speech recognition under limited or no supervision, including unsupervised, semi-supervised, and distant supervision settings.

Contribution

It introduces the largest open-source speech dataset for low-resource ASR and provides baseline systems and evaluation metrics for various limited supervision scenarios.

Findings

Largest freely-available speech corpus with 60K hours of audio

Baseline systems established for unsupervised, semi-supervised, and distant supervision

Evaluation metrics enable comparison with supervised state-of-the-art

Abstract

We introduce a new collection of spoken English audio suitable for training speech recognition systems under limited or no supervision. It is derived from open-source audio books from the LibriVox project. It contains over 60K hours of audio, which is, to our knowledge, the largest freely-available corpus of speech. The audio has been segmented using voice activity detection and is tagged with SNR, speaker ID and genre descriptions. Additionally, we provide baseline systems and evaluation metrics working under three settings: (1) the zero resource/unsupervised setting (ABX), (2) the semi-supervised setting (PER, CER) and (3) the distant supervision setting (WER). Settings (2) and (3) use limited textual resources (10 minutes to 10 hours) aligned with the speech. Setting (3) uses large amounts of unaligned text. They are evaluated on the standard LibriSpeech dev and test sets for comparison with the supervised state-of-the-art.