Whisper largeV2 Spanish MLS
This model is a fine-tuned version of openai/whisper-large-v2 on the facebook/multilingual_librispeech spanish dataset. It achieves the following results on the evaluation set:
- Loss: 0.0910
- Wer: 3.4678
Model description
The model is fine-tuned for 4000 updates/steps on multilingual librispeech Spanish train data.
- Zero-shot - 4.2 (MLS Spanish test)
- Fine-tune MLS spanish train - 3.46 (MLS Spanish test) (-17.61%)
- Zero-shot - 6.3 (CV11 test)
- Fine-tune MLS spanish train - 8.38 (CV11 test)
When the model is fine-tuned on specific dataset, the model loose its ability to generalise across datasets. Here the model is fine-tuned on MLS Spanish and evaluated on CV11 Spanish test. We can observe the drop in performance on CV11 test data.
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: 32
- eval_batch_size: 8
- seed: 42
- gradient_accumulation_steps: 2
- total_train_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: 4000
- mixed_precision_training: Native AMP
Training results
| Training Loss | Epoch | Step | Validation Loss | Wer |
|---|---|---|---|---|
| 0.2176 | 0.25 | 1000 | 0.1200 | 5.3932 |
| 0.1845 | 0.5 | 2000 | 0.1055 | 4.2 |
| 0.4516 | 0.75 | 3000 | 0.0977 | 3.6768 |
| 0.1549 | 1.14 | 4000 | 0.0910 | 3.4678 |
Framework versions
- Transformers 4.26.0.dev0
- Pytorch 1.13.0+cu117
- Datasets 2.7.1.dev0
- Tokenizers 0.13.2