This model is a fine-tuned version of cahya/wav2vec2-base-turkish-artificial on the MOZILLA-FOUNDATION/COMMON_VOICE_7_0 - TR dataset. It achieves the following results on the evaluation set:
- Loss: 0.2893
- Wer: 0.2713
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: 128
- eval_batch_size: 8
- seed: 42
- gradient_accumulation_steps: 4
- total_train_batch_size: 512
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- lr_scheduler_warmup_steps: 100
- num_epochs: 100.0
- mixed_precision_training: Native AMP
Training results
| Training Loss | Epoch | Step | Validation Loss | Wer |
|---|---|---|---|---|
| 1.8647 | 14.28 | 200 | 0.2758 | 0.2568 |
| 1.3376 | 28.56 | 400 | 0.2754 | 0.2722 |
| 1.1975 | 42.84 | 600 | 0.2929 | 0.2901 |
| 1.1024 | 57.14 | 800 | 0.2904 | 0.2928 |
| 1.0257 | 71.42 | 1000 | 0.2915 | 0.2823 |
| 0.9628 | 85.7 | 1200 | 0.2936 | 0.2749 |
| 0.9109 | 99.98 | 1400 | 0.2893 | 0.2713 |
Framework versions
- Transformers 4.17.0.dev0
- Pytorch 1.10.2+cu102
- Datasets 1.18.3
- Tokenizers 0.11.0