LibriSpeech test-clean
One of the most widely used public benchmark datasets in speech recognition, derived from LibriVox audiobook recordings, entirely in English at a 16 kHz sampling rate. The test-clean subset has been manually verified to remove samples with obvious noise or unclear pronunciation, and is commonly used to compare model performance under ideal conditions.
Timeline (last 90 days)
On the 50 samples of LibriSpeech test-clean, the WER gap between Parakeet and Whisper is 0.39 percentage points, which is on the same order of magnitude as the noise introduced by text normalization and cannot clearly distinguish which model is better.
LibriSpeech test-clean contains 40 speakers and 2620 utterances, from which the author drew 50 stratified samples (10 speakers, 5 utterances each), about 1002 reference words and 356 seconds of audio.
The author notes that this benchmark covers only English, clean read speech, a single M4 machine, and is a snapshot (the application has since been upgraded to FluidAudio 0.15.8, which fixed an issue where the endings of some long recordings were discarded).
All Facts (3)
On the 50 samples of LibriSpeech test-clean, the WER gap between Parakeet and Whisper is 0.39 percentage points, which is on the same order of magnitude as the noise introduced by text normalization and cannot clearly distinguish which model is better.
50%UnverifiedLibriSpeech test-clean contains 40 speakers and 2620 utterances, from which the author drew 50 stratified samples (10 speakers, 5 utterances each), about 1002 reference words and 356 seconds of audio.
50%UnverifiedThe author notes that this benchmark covers only English, clean read speech, a single M4 machine, and is a snapshot (the application has since been upgraded to FluidAudio 0.15.8, which fixed an issue where the endings of some long recordings were discarded).
50%