A repo for implementing popular ASR algorithms
| Model | WER(%) |
|---|---|
| Whisper large v3 turbo (baseline) | 16.22 |
| WavLM CTC loss | 11.72 |
WavLM CTC loss Command:
python run_wavlm.py --n_workers 16 --finetune_epoch 15 --run_name train_batches1.0 --val_wer --smoothing 0.1 --layer_weights --augment --augment_epoch 18 --logit_dropout_p 0.1 --finetune_lr_mult 1 --min_occurrences 15 --limit_train_batches 1.0 --limit_val_batches 1.0