Prosper700/naija-asr-backend

FastAPI backend for Nigerian language speech recognition and translation # Naija ASR & Translation API A FastAPI backend that powers speech recognition and translation for three Nigerian languages — **Hausa, Yoruba, and Igbo** — plus English. It transcribes voice to text with fine-tuned Whisper models and translates between languages with Meta's NLLB-200. Deployed on Hugging Face Spaces with a permanent public endpoint. > **Mobile app:** This backend serves the Naija Translator mobile app. See that repo for the React Native client. ## Features ### Speech Recognition (ASR) - **Nigerian language transcription** — Hausa, Yoruba, and Igbo via a fine-tuned Whisper model (`codeLumi/whisper-naija-v3b`, ~10% WER) - **English transcription** — high-accuracy English voice via `openai/whisper-base` - **Automatic model routing** — the API picks the right model based on the requested language - **Multi-format audio** — accepts m4a, mp3, wav, and webm (decoded via ffmpeg) ### Translation - **Bidirectional translation** — between English and any of the three Nigerian languages - **Nigerian-to-Nigerian** — e.g. Hausa to Yoruba - Powered by **NLLB-200** with FLORES language codes ### Pipeline - **One-shot voice translation** — audio in, translated text out (ASR + translation chained) ## Tech Stack | Layer | Technology | | ---------------- | -------------------------------------------- | | API framework | FastAPI | | ASR models | Whisper (fine-tuned + base) via Transformers | | Translation | NLLB-200 (distilled 1.3B) | | Audio processing | librosa, soundfile, ffmpeg | | ML runtime | PyTorch | | Server | Uvicorn | | Container | Docker | | Hosting | Hugging Face Spaces | ## Models All models are hosted on Hugging Face: | Model …