wav2vec2-xls-r-300m-Turkish-Tr-small
This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the common_voice dataset. It achieves the following results on the evaluation set:
- Loss: 0.4375
- Wer: 0.5050
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
- mixed_precision_training: Native AMP
Training results
| Training Loss | Epoch | Step | Validation Loss | Wer |
|---|---|---|---|---|
| 4.8735 | 4.21 | 400 | 2.8173 | 1.0002 |
| 1.0073 | 8.42 | 800 | 0.4981 | 0.6717 |
| 0.3395 | 12.63 | 1200 | 0.4470 | 0.5866 |
| 0.2254 | 16.84 | 1600 | 0.4349 | 0.5491 |
| 0.1648 | 21.05 | 2000 | 0.4454 | 0.5284 |
| 0.1325 | 25.26 | 2400 | 0.4552 | 0.5131 |
| 0.1102 | 29.47 | 2800 | 0.4375 | 0.5050 |
Framework versions
- Transformers 4.11.3
- Pytorch 1.10.0+cu111
- Datasets 1.14.0
- Tokenizers 0.10.3