wav2vec2-large-xls-r-300m-pa-in
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: 1.9680
- Wer: 0.7283
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: 16
- 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: 180
- mixed_precision_training: Native AMP
Training results
| Training Loss | Epoch | Step | Validation Loss | Wer |
|---|---|---|---|---|
| 8.2615 | 24.97 | 400 | 3.4784 | 1.0 |
| 3.366 | 49.97 | 800 | 2.3662 | 0.9917 |
| 1.1678 | 74.97 | 1200 | 1.4806 | 0.7709 |
| 0.5496 | 99.97 | 1600 | 1.7166 | 0.7476 |
| 0.4101 | 124.97 | 2000 | 1.8473 | 0.7510 |
| 0.3317 | 149.97 | 2400 | 1.9177 | 0.7322 |
| 0.2956 | 174.97 | 2800 | 1.9680 | 0.7283 |
Framework versions
- Transformers 4.16.2
- Pytorch 1.10.1+cu102
- Datasets 1.17.1.dev0
- Tokenizers 0.10.3
Evaluations Result
- WER: 0.7539
- CER: 0.2928