whisper-tiny-urdu
This model is a fine-tuned version of openai/whisper-tiny on the common_voice_17_0 dataset. It achieves the following results on the evaluation set:
- Loss: 0.7225
- Wer: 47.8529
Quick Usage
from transformers import pipeline
transcriber = pipeline(
"automatic-speech-recognition",
model="kingabzpro/whisper-tiny-urdu"
)
transcriber.model.generation_config.forced_decoder_ids = None
transcriber.model.generation_config.language = "ur"
transcription = transcriber("audio2.mp3")
print(transcription)
{'text': 'دیکھیے پانی کب تک بہتا اور مچھلی کب تک تیرتی ہے'}
Evaluation
| Dataset | WER (%) | CER (%) | BLEU | ChrF |
|---|---|---|---|---|
| Common Voice 17.0 (Urdu) | 46.908 | 18.543 | 32.631 | 63.988 |
| HowMannyMore/urdu-audiodataset | 51.405 | 21.830 | 31.475 | 64.204 |
Training procedure
Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 2e-05
- train_batch_size: 32
- eval_batch_size: 32
- seed: 42
- optimizer: Use OptimizerNames.ADAMW_TORCH with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
- lr_scheduler_type: cosine
- lr_scheduler_warmup_steps: 200
- training_steps: 2500
- mixed_precision_training: Native AMP
Training results
| Training Loss | Epoch | Step | Validation Loss | Wer |
|---|---|---|---|---|
| 0.6808 | 1.6949 | 500 | 0.7403 | 52.6699 |
| 0.3948 | 3.3898 | 1000 | 0.6850 | 47.1247 |
| 0.2873 | 5.0847 | 1500 | 0.6994 | 48.1516 |
| 0.2024 | 6.7797 | 2000 | 0.7169 | 46.7326 |
| 0.183 | 8.4746 | 2500 | 0.7225 | 47.8529 |
Framework versions
- Transformers 4.51.3
- Pytorch 2.6.0+cu124
- Datasets 3.6.0
- Tokenizers 0.21.1