Table 3. CER and relative CER reduction of various evaluation sets

Evaluation set Whisper model
large-v2 Model A Model B
KsponSpeech eval set 13.95 9.44 (32.33) 9.17 (34.26)
LibriSpeech test-clean 1.77 1.19 (32.77) 1.33 (24.86)
LibriSpeech test-other 2.86 2.87 (–0.35) 3.39 (–18.53)
CER, character error rate.