SKYLy
This model is a fine-tuned version of facebook/wav2vec2-large-xlsr-53 on the common_voice dataset. It achieves the following results on the evaluation set:
- Loss: 0.7645
- Wer: 0.4083
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: 0.0003
- train_batch_size: 16
- eval_batch_size: 8
- seed: 42
- gradient_accumulation_steps: 2
- total_train_batch_size: 32
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- lr_scheduler_warmup_steps: 500
- num_epochs: 30
Training results
| Training Loss | Epoch | Step | Validation Loss | Wer |
|---|---|---|---|---|
| 4.4215 | 4.26 | 400 | 1.6323 | 0.9857 |
| 0.5716 | 8.51 | 800 | 0.6679 | 0.5107 |
| 0.1721 | 12.77 | 1200 | 0.6935 | 0.4632 |
| 0.1063 | 17.02 | 1600 | 0.7533 | 0.4432 |
| 0.0785 | 21.28 | 2000 | 0.7208 | 0.4255 |
| 0.0608 | 25.53 | 2400 | 0.7481 | 0.4117 |
| 0.0493 | 29.79 | 2800 | 0.7645 | 0.4083 |
Framework versions
- Transformers 4.11.3
- Pytorch 1.10.0+cu113
- Datasets 2.1.0
- Tokenizers 0.10.3