Portuguese Medium Whisper
This model is a fine-tuned version of openai/whisper-medium on the common_voice_11_0 dataset. It achieves the following results on the evaluation set:
- Loss: 0.2628
- Wer: 6.5987
Blog post
All information about this model in this blog post: Speech-to-Text & IA | Transcreva qualquer áudio para o português com o Whisper (OpenAI)... sem nenhum custo!.
New SOTA
The Normalized WER in the OpenAI Whisper article with the Common Voice 9.0 test dataset is 8.1.
As this test dataset is similar to the Common Voice 11.0 test dataset used to evaluate our model (WER and WER Norm), it means that our Portuguese Medium Whisper is better than the Medium Whisper model at transcribing audios Portuguese in text (and even better than the Whisper Large that has a WER Norm of 7.1!).
Training procedure
Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 9e-06
- train_batch_size: 32
- eval_batch_size: 16
- seed: 42
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- lr_scheduler_warmup_steps: 500
- training_steps: 6000
- mixed_precision_training: Native AMP
Training results
| Training Loss | Epoch | Step | Validation Loss | Wer |
|---|---|---|---|---|
| 0.0333 | 2.07 | 1500 | 0.2073 | 6.9770 |
| 0.0061 | 5.05 | 3000 | 0.2628 | 6.5987 |
| 0.0007 | 8.03 | 4500 | 0.2960 | 6.6979 |
| 0.0004 | 11.0 | 6000 | 0.3212 | 6.6794 |
Framework versions
- Transformers 4.26.0.dev0
- Pytorch 1.13.0+cu117
- Datasets 2.7.1.dev0
- Tokenizers 0.13.2