LibriSpeech

A free dataset of roughly 1,000 hours of English read aloud from public-domain audiobooks recorded by LibriVox volunteers, cut into utterances, matched to the book text and stored at 16 kHz. It is split into training sets of 100 and 360 hours of 'clean' speech and 500 hours of 'other', more challenging speech, plus development and test sets of about five hours each, also divided into clean and other according to how hard a baseline recognizer found each speaker; no speaker appears in more than one split. A large text corpus for training language modelslanguage model, n.: A system that has learned how likely different sequences of words are, so it can predict the next word in a text and, by doing that repeatedly, write text of its own. is included. Released under a Creative Commons Attribution 4.0 license, it became the default benchmarkbenchmark, n.: A fixed, shared test — a set of questions or tasks with an agreed way of scoring — used to compare AI systems with one another and to track progress over time. for English speech recognitionspeech recognition, n.: Technology that turns spoken words into written text, as when a phone transcribes a voice message or captions appear automatically under a video., with results reported as word error rateword error rate, n.: The standard score for speech recognition: the share of words a system gets wrong compared with a correct transcript, counting wrong, missing and extra words. Lower is better. on test-clean and test-other. Best reported WERs fell below 2% on test-clean by the early 2020s, so the benchmark is near saturation and says little about conversational, noisy or accented speech. Derived resources include LibriTTS for speech synthesis and the 60,000-hour unlabeled LibriLight.

Good LibriSpeech numbers tell you little about our call-center audio, so we test on our own recordings too.