Model description
Chillarmo/whisper-small-hy-AM is an AI model designed for speech-to-text conversion specifically tailored for the Armenian language. Leveraging the power of fine-tuning, this model, named whisper-small-hy-AM, is based on openai/whisper-small and trained on the common_voice_16_1 dataset. It achieves the following results on the evaluation set:
- Loss: 0.2853
- Wer: 38.1160
Training Data and Future Enhancements
The training data consists of Mozilla Common Voice version 16.1. Plans for future improvements include continuing the training process and integrating an additional 10 hours of data from datasets such as google/fleurs and possibly google/xtreme_s. Despite its current performance, efforts are underway to further reduce the WER.
Training procedure
Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 1e-05
- train_batch_size: 16
- eval_batch_size: 8
- 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: 4000
- mixed_precision_training: Native AMP
Training results
| Training Loss | Epoch | Step | Validation Loss | Wer |
|---|---|---|---|---|
| 0.0989 | 2.48 | 1000 | 0.1948 | 41.5758 |
| 0.03 | 4.95 | 2000 | 0.2165 | 39.1251 |
| 0.0016 | 7.43 | 3000 | 0.2659 | 38.4089 |
| 0.0005 | 9.9 | 4000 | 0.2853 | 38.1160 |
Framework versions
- Transformers 4.37.2
- Pytorch 2.1.0+cu121
- Datasets 2.16.1
- Tokenizers 0.15.1