SpeechStew (1B)

Reported on 4 benchmarks across 1 task · 1 paper · 2 SOTA

Note: results are matched by exact model name. Different papers may use the same name for different model variants.

Audio4 results

Speech RecognitiononCHiME-6 eval
Word Error Rate (WER)· 2021-04-05
38.9
best: 30.7 (ConformerXXL-PS + G-Augment)
SOTA
SpeechStew: Simply Mix All Available Speech Recognition Data to Train One Large Neural Network arXiv:2104.02133
Speech RecognitiononCHiME-6 dev_gss12
Word Error Rate (WER)· 2021-04-05
31.9
best: 26 (ConformerXXL-PS + G-Augment)
SOTA
SpeechStew: Simply Mix All Available Speech Recognition Data to Train One Large Neural Network arXiv:2104.02133
Speech RecognitiononLibriSpeech test-clean
Word Error Rate (WER)· uses extra data· 2021-04-05
1.7
best: 0.985 (United Med ASR)
SpeechStew: Simply Mix All Available Speech Recognition Data to Train One Large Neural Network arXiv:2104.02133
Speech RecognitiononLibriSpeech test-other
Word Error Rate (WER)· 2021-04-05
3.3
best: 2.48 (SAMBA ASR)
SpeechStew: Simply Mix All Available Speech Recognition Data to Train One Large Neural Network arXiv:2104.02133