<!-- This model card has been generated automatically according to the information the Trainer had access to. You should probably proofread and complete it, then remove this comment. -->
wav2vec2-large-xlsr-53-demo-colab
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: 0.6901
- Wer: 1.6299
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: 30
- mixed_precision_training: Native AMP
Training results
| Training Loss | Epoch | Step | Validation Loss | Wer |
|---|---|---|---|---|
| 8.5034 | 3.42 | 400 | 3.5852 | 1.0 |
| 1.7853 | 6.83 | 800 | 0.7430 | 1.6774 |
| 0.5675 | 10.26 | 1200 | 0.6513 | 1.6330 |
| 0.3761 | 13.67 | 1600 | 0.6208 | 1.6081 |
| 0.2776 | 17.09 | 2000 | 0.6401 | 1.6081 |
| 0.2266 | 20.51 | 2400 | 0.6410 | 1.6295 |
| 0.1949 | 23.93 | 2800 | 0.6910 | 1.6287 |
| 0.1672 | 27.35 | 3200 | 0.6901 | 1.6299 |
Framework versions
- Transformers 4.11.3
- Pytorch 1.10.0+cu111
- Datasets 1.14.0
- Tokenizers 0.10.3
