Tamziycode/speak-nigeria-now

# Multilingual STT Prototype A research-grade speech-to-text web application for **Yoruba**, **Hausa**, and **Nigerian Pidgin**, built to evaluate state-of-the-art multilingual ASR performance on low-resource Nigerian languages. Audio captured in-browser (or uploaded) is normalized to 16 kHz mono WAV and proxied through a Cloudflare Worker to the Groq API, which serves OpenAI's `whisper-large-v3` model at temperature `0.0` for deterministic transcription. Transcripts can be edited, copied, exported to `.txt`, and translated to English on demand. ## Technical Stack | Layer | Technology | | ----------- | ----------------------------------------------------------- | | Frontend | React 19, TanStack Start, TanStack Router, Tailwind CSS v4 | | Build | Vite 7 | | Runtime | Cloudflare Workers (serverless edge) | | Audio DSP | Web Audio API (`AudioContext`, `OfflineAudioContext`) | | ASR Model | `whisper-large-v3` via Groq Inference API | | Persistence | `localStorage` (last session recovery) | ## Core Features - Live microphone capture with WAV normalization (16 kHz mono PCM) - Drag-and-drop / file upload (`.wav`, `.mp3`, `.m4a`, max 25 MB) - Editable transcript with real-time word and character analytics - One-click English translation via Whisper's translation endpoint - Copy to clipboard and `.txt` export for both transcript and translation - Session recovery across page reloads - Live API health indicator, mic-permission diagnostics, offline detection ## Supported Languages | Language | ISO Code | Strategy | | --------------- | -------- | --------------------------------------------------- | | Hausa | `ha` | Native Whisper language hint | | Yoruba | `yo` | Native Whisper la …