wav2vec2-base-timit-demo-colab-32
This model is a fine-tuned version of facebook/wav2vec2-base on an unknown dataset. It achieves the following results on the evaluation set:
- Loss: 0.4488
- Wer: 0.3149
Model description
More information needed
Intended uses & limitations
More information needed
Training and evaluation data
More information needed
Training procedure
Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 0.0001
- train_batch_size: 32
- eval_batch_size: 8
- seed: 42
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- lr_scheduler_warmup_steps: 1000
- num_epochs: 30
- mixed_precision_training: Native AMP
Training results
| Training Loss | Epoch | Step | Validation Loss | Wer |
|---|---|---|---|---|
| 3.6155 | 4.0 | 500 | 2.2647 | 0.9992 |
| 0.9037 | 8.0 | 1000 | 0.4701 | 0.4336 |
| 0.3159 | 12.0 | 1500 | 0.4247 | 0.3575 |
| 0.1877 | 16.0 | 2000 | 0.4477 | 0.3442 |
| 0.1368 | 20.0 | 2500 | 0.4932 | 0.3384 |
| 0.1062 | 24.0 | 3000 | 0.4758 | 0.3202 |
| 0.0928 | 28.0 | 3500 | 0.4488 | 0.3149 |
Framework versions
- Transformers 4.12.5
- Pytorch 1.10.0+cu111
- Tokenizers 0.10.3