wav2vec2-xlsr-ft-cy-verbatim
This model is a fine-tuned version of facebook/wav2vec2-large-xlsr-53 on the techiaith/banc-trawsgrifiadau-bangor dataset. It achieves the following results on the evaluation set:
- Loss: 0.4357
- Wer: 0.3262
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: 5.0
- mixed_precision_training: Native AMP
Training results
| Training Loss | Epoch | Step | Validation Loss | Wer |
|---|---|---|---|---|
| No log | 0.21 | 100 | 3.4135 | 1.0 |
| No log | 0.41 | 200 | 2.9521 | 1.0 |
| No log | 0.62 | 300 | 2.3339 | 0.9365 |
| No log | 0.83 | 400 | 1.2433 | 0.8259 |
| 3.1912 | 1.03 | 500 | 0.8614 | 0.6385 |
| 3.1912 | 1.24 | 600 | 0.7557 | 0.5612 |
| 3.1912 | 1.44 | 700 | 0.6781 | 0.5195 |
| 3.1912 | 1.65 | 800 | 0.6363 | 0.4879 |
| 3.1912 | 1.86 | 900 | 0.5959 | 0.4559 |
| 0.8237 | 2.06 | 1000 | 0.5430 | 0.4260 |
| 0.8237 | 2.27 | 1100 | 0.5293 | 0.4098 |
| 0.8237 | 2.48 | 1200 | 0.5141 | 0.4056 |
| 0.8237 | 2.68 | 1300 | 0.4879 | 0.3947 |
| 0.8237 | 2.89 | 1400 | 0.4697 | 0.3788 |
| 0.5625 | 3.1 | 1500 | 0.4748 | 0.3780 |
| 0.5625 | 3.3 | 1600 | 0.4836 | 0.3684 |
| 0.5625 | 3.51 | 1700 | 0.4796 | 0.3625 |
| 0.5625 | 3.72 | 1800 | 0.4582 | 0.3515 |
| 0.5625 | 3.92 | 1900 | 0.4395 | 0.3437 |
| 0.4267 | 4.13 | 2000 | 0.4410 | 0.3420 |
| 0.4267 | 4.33 | 2100 | 0.4467 | 0.3382 |
| 0.4267 | 4.54 | 2200 | 0.4398 | 0.3329 |
| 0.4267 | 4.75 | 2300 | 0.4383 | 0.3287 |
| 0.4267 | 4.95 | 2400 | 0.4358 | 0.3264 |
Framework versions
- Transformers 4.28.1
- Pytorch 2.0.0+cu117
- Datasets 2.11.0
- Tokenizers 0.13.3