Speech Recognition on Librispeech (100 hours) and AISHELL v1
5.4WER (English)FERERO-FT
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| FERERO-FTfine-tuning_mode=FERERO, losses=CTC losses + CPC loss2024.12 | 5.4 | 4.9 | 5.1 | |
| LS-FTfine-tuning_mode=Linear Scalarization, losses=CTC losses + CPC loss2024.12 | 6.8 | 5.9 | 6.4 | |
| Komatsu et al.2024.12 | 7.11 | — | — | |
| Init. (M2ASR)model_state=Initial pre-trained2024.12 | 7.3 | 6.2 | 6.7 | |
| w/o CPCoptimization_strategy=Sum of supervised CTC losses2024.12 | 11.8 | 10.2 | 11 |