Data-Efficient On-Policy Distillation for Automatic Speech Recognition

Benchmark Model Rank Results
speech-recognition-on-gigaspeech-cleanedAutoArk-AI/ARK-ASR-3B#1Word Error Rate (WER): 6.94
speech-recognition-on-gigaspeech-cleanedAutoArk-AI/ARK-ASR-0.6B#4Word Error Rate (WER): 7.38
speech-recognition-on-gigaspeech-open-asr-leaderboardArk-ASR#1WER: 8.0
speech-recognition-on-gigaspeech-testAutoArk-AI/ARK-ASR-3B#1Word Error Rate (WER): 6.98
speech-recognition-on-gigaspeech-testAutoArk-AI/ARK-ASR-0.6B#4Word Error Rate (WER): 7.42
speech-recognition-on-ted-lium-open-asr-leaderboardArk-ASR#2WER: 2.45
speech-recognition-on-earnings-22AutoArk-AI/ARK-ASR-3B#3Word Error Rate (WER): 8.23
speech-recognition-on-earnings-22AutoArk-AI/ARK-ASR-0.6B#9Word Error Rate (WER): 9.69
speech-recognition-on-librispeech-test-cleanAutoArk-AI/ARK-ASR-3B#3Word Error Rate (WER): 1.03
speech-recognition-on-librispeech-test-cleanAutoArk-AI/ARK-ASR-0.6B#33Word Error Rate (WER): 1.47
speech-recognition-on-librispeech-test-otherAutoArk-AI/ARK-ASR-3B#5Word Error Rate (WER): 2.35
speech-recognition-on-librispeech-test-otherAutoArk-AI/ARK-ASR-0.6B#47Word Error Rate (WER): 3.44
speech-recognition-on-ami-ihm-open-asr-leaderboardArk-ASR#6WER: 10.02
speech-recognition-on-voxpopuliAutoArk-AI/ARK-ASR-3B#7Word Error Rate (WER): 5.47
speech-recognition-on-voxpopuliAutoArk-AI/ARK-ASR-0.6B#23Word Error Rate (WER): 6.29
speech-recognition-on-amiAutoArk-AI/ARK-ASR-3B#8Word Error Rate (WER): 8.79
speech-recognition-on-amiAutoArk-AI/ARK-ASR-0.6B#13Word Error Rate (WER): 9.81
speech-recognition-on-ami-cleanedAutoArk-AI/ARK-ASR-3B#8Word Error Rate (WER): 7.89
speech-recognition-on-ami-cleanedAutoArk-AI/ARK-ASR-0.6B#12Word Error Rate (WER): 8.6
speech-recognition-on-voxpopuli-cleanedAutoArk-AI/ARK-ASR-0.6B#10Word Error Rate (WER): 2.76
speech-recognition-on-voxpopuli-cleanedAutoArk-AI/ARK-ASR-3B#22Word Error Rate (WER): 3.55
speech-recognition-on-voxpopuli-open-asr-leaderboardArk-ASR#13WER: 6.31
speech-recognition-on-spgispeechAutoArk-AI/ARK-ASR-3B#22Word Error Rate (WER): 2.46
speech-recognition-on-spgispeechAutoArk-AI/ARK-ASR-0.6B#29Word Error Rate (WER): 2.63