Captions for Roman Urdu speakers
CutPro · Evolvo (QuantumBytes)
- Problem:
- Creators in Pakistan write Urdu in English letters, but no speech-to-text engine outputs Roman Urdu with word timings. Whisper, AssemblyAI, Qwen ASR and a fine-tuned model all fell short.
- What I built:
- A pipeline that transcribes Urdu script, transliterates it (neural model plus a Rust/WASM rule-based fallback), keeps English words in English and repairs word timing and repetition loops.
- Result:
- Word-by-word Roman Urdu and Roman Hindi captions in CutPro. On my test sets: rule-based accuracy from 66.7% to 95.8%, timing error from 279 ms to 162 ms.