FairASR: Fair Audio Contrastive Learning for Automatic Speech Recognition
Paper
•
2506.10747
•
Published
audio
audioduration (s) 1.36
65.6
|
|---|
Fair-Speech dataset, a public corpus containing 26,500 utterances from 593 compensated U.S. participants with self-reported demographics including age, gender, ethnicity, location, and native language status.
The dataset spans seven voice assistant domains: music, capture, utilities, notification control, messaging, calling, and dictation.
Total number of unique speakers: 593
Total number of utterances: 26,471
Average per utterance length: 7.42 seconds
100% US (excludes speakers in IL, WA and TX)
FairASR: Fair Audio Contrastive Learning for Automatic Speech Recognition