Whisper Small Hy - Erik Mkrtchyan
This model is a fine-tuned version of openai/whisper-small on the Hy Generated Audio Data with CV 20.0 dataset. It achieves the following results on the evaluation set:
- Loss: 0.1685
- Wer: 36.7015
Model description
This model is based on OpenAI's Whisper Small and fine-tuned for Armenian using exclusively real audio data. It is designed to transcribe Armenian speech into text and serves as a benchmark to evaluate how well the model learns using only real (non-synthetic) data.
Training and evaluation data
The dataset contains both real and high-quality synthetic Armenian speech clips.
| Split | # Clips | Duration (hours) |
|---|---|---|
train |
9,300 | 13.53 |
test |
5,818 | 9.16 |
eval |
5,856 | 8.76 |
Total duration: ~31 hours
Train set duration(train+generated): ~13 hours
Test set duration(test+eval) ~18 hours
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: 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: 500
- num_epochs: 3
- mixed_precision_training: Native AMP
Training results
| Training Loss | Epoch | Step | Validation Loss | Wer |
|---|---|---|---|---|
| 0.1946 | 0.9983 | 581 | 0.2224 | 48.2229 |
| 0.1212 | 1.9966 | 1162 | 0.1735 | 39.2161 |
| 0.077 | 2.9948 | 1743 | 0.1685 | 36.7015 |
Framework versions
- Transformers 4.51.3
- Pytorch 2.7.0+cu126
- Datasets 3.6.0
- Tokenizers 0.21.1