The model is a fine-tuned version of Meta model facebook/wav2vec2-large-lv60 on timit dataset. It yields aLos result the following:
WER: 0.247330
Validation Loss: 0.299390
The intention behind is to make the model Multilingual although yet it is just for english
Training result
| Step | Training Loss | Validation Loss | Wer |
|---|---|---|---|
| 500 | 4.808300 | 2.929884 | 1.000000 |
| 1000 | 1.543200 | 0.367235 | 0.415202 |
| 1500 | 0.305000 | 0.264133 | 0.321274 |
| 2000 | 0.161400 | 0.268306 | 0.294949 |
| 2500 | 0.109600 | 0.258425 | 0.276135 |
| 3000 | 0.084900 | 0.263341 | 0.268348 |
| 3500 | 0.066600 | 0.258404 | 0.259734 |
| 4000 | 0.055800 | 0.257602 | 0.249397 |
| 4500 | 0.045500 | 0.271701 | 0.253118 |
| 5000 | 0.042200 | 0.292905 | 0.251327 |
| 5500 | 0.038300 | 0.278083 | 0.248019 |
| 6000 | 0.033000 | 0.295660 | 0.250362 |
| 6500 | 0.031400 | 0.286542 | 0.248639 |
| 7000 | 0.030800 | 0.299390 | 0.247330 |
Training Procedure
Hyper parameters
More information to come
Tensorboard
More information to come
- Downloads last month
- 2