whisper-small-bem-v1
This model is a fine-tuned version of openai/whisper-small on an unknown dataset. It achieves the following results on the evaluation set:
- Loss: 0.4645
- Wer: 36.1826
When combined with fine-tuned NLLB-200 3.3B, Bemba-English results are as follows:
| BLEU | ChrF++ | COMET |
|---|---|---|
| 27.41 | 49.65 | 51.77 |
Model description
Bemba automatic speech recognition (ASR)
Intended uses & limitations
For research purposes only
Training and evaluation data
Big-C
Training procedure
Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 0.0001
- train_batch_size: 8
- eval_batch_size: 8
- seed: 42
- optimizer: Use OptimizerNames.ADAMW_TORCH with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
- lr_scheduler_type: linear
- lr_scheduler_warmup_ratio: 0.03
- num_epochs: 3
- mixed_precision_training: Native AMP
Training results
| Training Loss | Epoch | Step | Validation Loss | Wer |
|---|---|---|---|---|
| 0.5933 | 1.0 | 10645 | 0.5593 | 44.3113 |
| 0.4243 | 2.0 | 21290 | 0.4738 | 38.8064 |
| 0.2944 | 3.0 | 31935 | 0.4645 | 36.1826 |
Framework versions
- Transformers 4.51.3
- Pytorch 2.4.1+cu124
- Datasets 3.5.1
- Tokenizers 0.21.1