wav2vec2-large-xls-r-300m-hi
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: 2.4749
- Wer: 0.9420
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: 7.5e-05
- 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: 50
- mixed_precision_training: Native AMP
Training results
| Training Loss | Epoch | Step | Validation Loss | Wer |
|---|---|---|---|---|
| 9.8626 | 4.76 | 400 | 3.6151 | 1.0 |
| 3.5463 | 9.52 | 800 | 3.5778 | 1.0 |
| 3.4415 | 14.28 | 1200 | 3.4525 | 1.0 |
| 3.0927 | 19.05 | 1600 | 2.6220 | 0.9860 |
| 2.0573 | 23.8 | 2000 | 2.3974 | 0.9610 |
| 1.5905 | 28.57 | 2400 | 2.4427 | 0.9558 |
| 1.426 | 33.33 | 2800 | 2.4736 | 0.9475 |
| 1.3147 | 38.09 | 3200 | 2.4494 | 0.9417 |
| 1.2642 | 42.85 | 3600 | 2.4665 | 0.9450 |
| 1.2289 | 47.62 | 4000 | 2.4749 | 0.9420 |
Framework versions
- Transformers 4.11.3
- Pytorch 1.10.1+cu102
- Datasets 1.17.1.dev0
- Tokenizers 0.10.3