Urdu Speech Recognition

Fine-tuning OpenAI Whisper on Urdu speech data from Mozilla Common Voice.

Why

Urdu speech recognition is underserved compared to English and other high-resource languages. Whisper's multilingual pretraining gives a strong starting point — fine-tuning on Urdu-specific data should meaningfully reduce error rates.

Status

Planned — Year 3 of my BS.

Planned approach

• Use Mozilla Common Voice Urdu (100+ hours of labeled speech)
• Fine-tune whisper-small using HuggingFace's Seq2SeqTrainer
• Measure Word Error Rate (WER) before and after fine-tuning
• Deploy a working demo with Gradio on HuggingFace Spaces

Tech stack

Python, PyTorch, HuggingFace Transformers + Datasets, Gradio

Author

Salik Hussain — github.com/salikhussain71-code

License

MIT