wav2vec2-xls-r-300m-W2V2-XLSR-300M-YAKUT-SMALL
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: 0.9068
- Wer: 0.7900
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: 50
- mixed_precision_training: Native AMP
Training results
| Training Loss | Epoch | Step | Validation Loss | Wer |
|---|---|---|---|---|
| 4.6926 | 19.05 | 400 | 2.7538 | 1.0 |
| 0.7031 | 38.1 | 800 | 0.9068 | 0.7900 |
Framework versions
- Transformers 4.11.3
- Pytorch 1.10.0+cu111
- Datasets 1.14.0
- Tokenizers 0.10.3
- Downloads last month
- 11
Dataset used to train emre/wav2vec2-xls-r-300m-W2V2-XLSR-300M-YAKUT-SMALL
Evaluation results
- Test WER on Common Voice sahself-reported79.000