This model is a fine-tuned version of KBLab/wav2vec2-large-voxrex on the MOZILLA-FOUNDATION/COMMON_VOICE_9_0 - SV-SE dataset. It achieves the following results on the evaluation set ("test" split, without LM):
- Loss: 0.1318
- Wer: 0.1121
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: 7.5e-05
- train_batch_size: 32
- eval_batch_size: 32
- seed: 42
- gradient_accumulation_steps: 4
- total_train_batch_size: 128
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- lr_scheduler_warmup_ratio: 0.2
- num_epochs: 100.0
- mixed_precision_training: Native AMP
Training results
| Training Loss | Epoch | Step | Validation Loss | Wer |
|---|---|---|---|---|
| 2.9099 | 10.42 | 1000 | 2.8369 | 1.0 |
| 1.0745 | 20.83 | 2000 | 0.1957 | 0.1673 |
| 0.934 | 31.25 | 3000 | 0.1579 | 0.1389 |
| 0.8691 | 41.66 | 4000 | 0.1457 | 0.1290 |
| 0.8328 | 52.08 | 5000 | 0.1435 | 0.1205 |
| 0.8068 | 62.5 | 6000 | 0.1350 | 0.1191 |
| 0.7822 | 72.91 | 7000 | 0.1347 | 0.1155 |
| 0.7769 | 83.33 | 8000 | 0.1321 | 0.1131 |
| 0.7678 | 93.75 | 9000 | 0.1321 | 0.1115 |
Framework versions
- Transformers 4.17.0.dev0
- Pytorch 1.10.2+cu102
- Datasets 2.2.2
- Tokenizers 0.11.0