wav2vec2-large-xls-r-300m-lithuanian
This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the MOZILLA-FOUNDATION/COMMON_VOICE_7_0 - LT dataset. It achieves the following results on the evaluation set:
- Loss: 0.1722
- Wer: 0.2486
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: 7e-05
- train_batch_size: 32
- eval_batch_size: 1
- seed: 42
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- lr_scheduler_warmup_steps: 2000
- num_epochs: 50.0
- mixed_precision_training: Native AMP
Training results
| Training Loss | Epoch | Step | Validation Loss | Wer |
|---|---|---|---|---|
| 1.6837 | 8.0 | 2000 | 0.6649 | 0.7515 |
| 1.1105 | 16.0 | 4000 | 0.2386 | 0.3436 |
| 1.0069 | 24.0 | 6000 | 0.2008 | 0.2968 |
| 0.9417 | 32.0 | 8000 | 0.1915 | 0.2774 |
| 0.887 | 40.0 | 10000 | 0.1819 | 0.2616 |
| 0.8563 | 48.0 | 12000 | 0.1729 | 0.2475 |
Framework versions
- Transformers 4.16.0.dev0
- Pytorch 1.10.1+cu102
- Datasets 1.17.1.dev0
- Tokenizers 0.11.0