NepaliConformer — Nepali ASR for real telephone calls

121M Conformer trained on ~1,655 h of conversational Nepali. 33.8% WER on real call-center audio (NepTel benchmark) where Whisper-large-v3 zero-shot scores ~99%. Honest limitations and the full benchmark: github.com/Ampixa/nepaliconformer. CPU demo — a 30 s clip takes roughly 10-20 s. Example clips are real call-center audio (CC-BY-4.0, © InfoBayAI).

Demo stuck or mic blocked? Open the demo full-screen · Model downloads and usage: github.com/Ampixa/nepaliconformer