whisper-zulu-medium
This model is a fine-tuned version of openai/whisper-medium on the wjbmattingly/zulu_merged_audio dataset. It achieves the following results on the evaluation set:
- Loss: 0.2949
- Wer: 0.1993
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: 1e-05
- train_batch_size: 16
- eval_batch_size: 16
- seed: 42
- distributed_type: multi-GPU
- num_devices: 4
- total_train_batch_size: 64
- total_eval_batch_size: 64
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- lr_scheduler_warmup_steps: 500
- training_steps: 400
- mixed_precision_training: Native AMP
Training results
| Training Loss | Epoch | Step | Validation Loss | Wer |
|---|---|---|---|---|
| 0.8378 | 1.25 | 100 | 0.7290 | 0.5605 |
| 0.3624 | 2.5 | 200 | 0.4048 | 0.2791 |
| 0.2279 | 3.75 | 300 | 0.3236 | 0.2187 |
| 0.1524 | 5.0 | 400 | 0.2949 | 0.1993 |
Framework versions
- Transformers 4.45.0.dev0
- Pytorch 2.0.1+cu117
- Datasets 2.21.0
- Tokenizers 0.19.1