wav2vec2-large-xls-r-300m-hausa
This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the MOZILLA-FOUNDATION/COMMON_VOICE_7_0 - HA dataset. It achieves the following results on the evaluation set:
- Loss: 0.5756
- Wer: 0.6014
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: 32
- seed: 42
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- lr_scheduler_warmup_steps: 500
- num_epochs: 100.0
- mixed_precision_training: Native AMP
Training results
| Training Loss | Epoch | Step | Validation Loss | Wer |
|---|---|---|---|---|
| 2.7064 | 11.36 | 500 | 2.7112 | 1.0 |
| 1.3079 | 22.73 | 1000 | 0.7337 | 0.7776 |
| 1.0919 | 34.09 | 1500 | 0.5938 | 0.7023 |
| 0.9546 | 45.45 | 2000 | 0.5698 | 0.6133 |
| 0.8895 | 56.82 | 2500 | 0.5739 | 0.6142 |
| 0.8152 | 68.18 | 3000 | 0.5579 | 0.6091 |
| 0.7703 | 79.55 | 3500 | 0.5813 | 0.6210 |
| 0.732 | 90.91 | 4000 | 0.5756 | 0.5860 |
Framework versions
- Transformers 4.16.0.dev0
- Pytorch 1.10.1+cu102
- Datasets 1.17.1.dev0
- Tokenizers 0.11.0