from transformers import WhisperForConditionalGeneration, WhisperProcessor
import torch
import librosa

# Load model
model_name = "servinosmanov/whisper-large-v3-crh"
processor = WhisperProcessor.from_pretrained(model_name)
model = WhisperForConditionalGeneration.from_pretrained(model_name)

device = "cuda" if torch.cuda.is_available() else "cpu"
model = model.to(device)
model.eval()

# Load and process audio
audio, sr = librosa.load("your_audio.wav", sr=16000)

input_features = processor(
    audio,
    sampling_rate=16000,
    return_tensors="pt"
).input_features.to(device)

# Generate transcription (with repetition penalty recommended)
with torch.no_grad():
    predicted_ids = model.generate(
        input_features,
        max_length=225,
        num_beams=5,
        repetition_penalty=1.2,
        no_repeat_ngram_size=3,
    )

transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)[0]
print(transcription)

Important Notes

Repetition Penalty: It's recommended to use repetition_penalty=1.2 and no_repeat_ngram_size=3 during inference to prevent repetition loops
Script Support: Works with both Cyrillic (Меним адым...) and Latin (Menim adım...) scripts

Comparison with whisper-medium

Model	WER
whisper-medium (fine-tuned)	26.47%
whisper-large-v3 (this model)	13.67%
Improvement	48.4% relative

Limitations

Trained primarily on female voice (Sevil speaker)
May have reduced accuracy on very long utterances
Best results with clear audio at 16kHz sampling rate

License

Apache 2.0 (same as base Whisper model)

Citation

If you use this model, please cite:

bibtex

@misc{whisper-large-v3-crh,
  author = {Servin Osmanov},
  title = {Whisper Large-v3 Fine-tuned for Crimean Tatar},
  year = {2025},
  publisher = {HuggingFace},
  url = {https://huggingface.co/servinosmanov/whisper-large-v3-crh}
}