whisper small finetuned TLT non-native child speech
This model is a fine-tuned version of openai/whisper-small on the LTL2021 dataset. It achieves the following results on the evaluation set:
- Loss: 0.4034
- Wer: 18.5163
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: 32
- eval_batch_size: 8
- seed: 42
- gradient_accumulation_steps: 2
- total_train_batch_size: 64
- optimizer: Use 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_steps: 2048
- training_steps: 4000
- mixed_precision_training: Native AMP
Training results
| Training Loss | Epoch | Step | Validation Loss | Wer |
|---|---|---|---|---|
| 5.0933 | 1.7857 | 500 | 2.9951 | 17.7932 |
| 3.1499 | 3.5714 | 1000 | 1.8692 | 16.9543 |
| 1.7621 | 5.3571 | 1500 | 1.0117 | 17.4699 |
| 0.7934 | 7.1429 | 2000 | 0.4569 | 19.5452 |
| 0.7021 | 8.9286 | 2500 | 0.4180 | 18.7828 |
| 0.6198 | 10.7143 | 3000 | 0.4087 | 19.0821 |
| 0.5799 | 12.5 | 3500 | 0.4047 | 19.8991 |
| 0.5980 | 14.2857 | 4000 | 0.4034 | 18.5163 |
Framework versions
- Transformers 5.8.1
- Pytorch 2.5.1+cu121
- Datasets 4.8.5
- Tokenizers 0.22.2