wav2vec2-large-xlsr-hindi_commonvoice
This model is a fine-tuned version of facebook/wav2vec2-large-xlsr-53 on the common_voice dataset. It achieves the following results on the evaluation set:
- Loss: 3.5947
- Wer: 1.0
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: 32
- eval_batch_size: 8
- seed: 42
- gradient_accumulation_steps: 2
- total_train_batch_size: 64
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- lr_scheduler_warmup_steps: 100
- num_epochs: 30
- mixed_precision_training: Native AMP
Training results
| Training Loss | Epoch | Step | Validation Loss | Wer |
|---|---|---|---|---|
| 24.0069 | 4.0 | 20 | 40.3956 | 1.0 |
| 18.1097 | 8.0 | 40 | 15.3603 | 1.0 |
| 7.1344 | 12.0 | 60 | 5.2695 | 1.0 |
| 4.0032 | 16.0 | 80 | 3.7403 | 1.0 |
| 3.4894 | 20.0 | 100 | 3.5724 | 1.0 |
| 3.458 | 24.0 | 120 | 3.6164 | 1.0 |
| 3.4412 | 28.0 | 140 | 3.5947 | 1.0 |
Framework versions
- Transformers 4.11.3
- Pytorch 1.10.0+cu102
- Datasets 1.13.3
- Tokenizers 0.10.3