HuBERT: Self-Supervised Speech Representation Learning by Masked Prediction of Hidden Units

Benchmark Model Rank Results
speech-recognition-on-amifacebook/hubert-xlarge-ls960-ft#58Word Error Rate (WER): 31.53
speech-recognition-on-amifacebook/hubert-large-ls960-ft#60Word Error Rate (WER): 32.32
speech-recognition-on-ami-cleanedfacebook/hubert-xlarge-ls960-ft#61Word Error Rate (WER): 28.15
speech-recognition-on-ami-cleanedfacebook/hubert-large-ls960-ft#63Word Error Rate (WER): 28.9
speech-recognition-on-earnings-22facebook/hubert-large-ls960-ft#61Word Error Rate (WER): 26.87
speech-recognition-on-earnings-22facebook/hubert-xlarge-ls960-ft#63Word Error Rate (WER): 27.25
speech-recognition-on-gigaspeech-cleanedfacebook/hubert-xlarge-ls960-ft#61Word Error Rate (WER): 15.77
speech-recognition-on-gigaspeech-cleanedfacebook/hubert-large-ls960-ft#63Word Error Rate (WER): 16.02
speech-recognition-on-gigaspeech-testfacebook/hubert-xlarge-ls960-ft#62Word Error Rate (WER): 16.01
speech-recognition-on-gigaspeech-testfacebook/hubert-large-ls960-ft#65Word Error Rate (WER): 16.28
speech-recognition-on-librispeech-test-cleanfacebook/hubert-xlarge-ls960-ft#33Word Error Rate (WER): 1.5
speech-recognition-on-librispeech-test-cleanfacebook/hubert-large-ls960-ft#38Word Error Rate (WER): 1.55
speech-recognition-on-librispeech-test-cleanHuBERT with Libri-Light#55Word Error Rate (WER): 1.8
speech-recognition-on-librispeech-test-otherHuBERT with Libri-Light#22Word Error Rate (WER): 2.9
speech-recognition-on-librispeech-test-otherfacebook/hubert-xlarge-ls960-ft#25Word Error Rate (WER): 3.02
speech-recognition-on-librispeech-test-otherfacebook/hubert-large-ls960-ft#46Word Error Rate (WER): 3.65
speech-recognition-on-spgispeechfacebook/hubert-xlarge-ls960-ft#65Word Error Rate (WER): 8.66
speech-recognition-on-spgispeechfacebook/hubert-large-ls960-ft#66Word Error Rate (WER): 8.95
speech-recognition-on-voxpopulifacebook/hubert-xlarge-ls960-ft#62Word Error Rate (WER): 14.99
speech-recognition-on-voxpopulifacebook/hubert-large-ls960-ft#63Word Error Rate (WER): 15.36
speech-recognition-on-voxpopuli-cleanedfacebook/hubert-xlarge-ls960-ft#63Word Error Rate (WER): 7.33
speech-recognition-on-voxpopuli-cleanedfacebook/hubert-large-ls960-ft#64Word Error Rate (WER): 7.78