Note About 1,000 hours of read English audiobooks; the standard ASR benchmark.

Note Very large open English speech corpus.

Note 10,000 hours of English audiobooks, podcasts and YouTube. Gated.

Note Read audiobook speech at scale, several languages.

Note English portion of Multilingual LibriSpeech.

Note European Parliament speech, multilingual.

Note Read-speech benchmark across about 100 languages.

Note YouTube-derived multilingual speech at very large scale.

Note Large multilingual ASR and translation corpus with pseudo-labels.

Note Meeting recordings: multi-speaker, conversational.

Note 5,000 hours of earnings calls, formatted transcripts. Gated.

Note Accented earnings-call speech; long-form benchmark.

Note Classic single-speaker TTS set. Older script loader.

Note LibriSpeech re-cut for TTS.

Note LibriTTS with restored audio quality.

Note About 110 English speakers, varied accents. Older script loader.

Note Large in-the-wild multilingual speech for TTS. Gated.

Note Short spoken keywords, for keyword spotting. Older script loader.

Note Multilingual spoken keywords.

Note Sound-event classification, not transcription.

Note Very large English speech corpus; newer GigaSpeech release.