pytomb/ghana-speech-bridge

Ghana language + voice bridge — Twi, Ewe, Ga, Dagbani. Speech API, i18n packs, and ML notebooks for fine-tuning Whisper/Gemma. # Ghana Speech Bridge > Voice-native language infrastructure for Ghanaian applications. > Supports Twi, Ewe, Ga, Dagbani, Fante — with code-switching and offline-first design.   --- ## What This Is A standalone service + ML workspace that gives any app voice and language capabilities for Ghanaian users — especially those without formal education who need to speak and hear in their native tongue. **Two parallel tracks:** | Track | What | Timeline | |-------|------|----------| | **Short-term** | API-powered bridge (Khaya AI) with smart caching + call budget | Now | | **Long-term** | Own every layer: fine-tuned Whisper ASR, Gemma 2B translation, Coqui TTS | Ongoing | The service interface stays constant — backend providers swap underneath. Every app built on this bridge gets improvements automatically. --- ## Architecture ``` User speaks (Twi / Ewe / Ga / Dagbani / mixed) │ ▼ [1] Language Detector → fast classifier, runs locally │ ▼ [2] ASR /api/stt → Khaya STT (now) → Whisper-Akan (soon) │ "Me pɛ sɛ me kɔ fie" ▼ [3] Translate /api/translate → Khaya API (now) → Gemma 2B fine-tuned (soon) │ "I want to go home" ▼ [4] LLM Reasoning → Claude / Gemini (always English, always cloud) │ "Here are 3 bus routes..." ▼ [5] Translate back → same as step 3 │ ▼ [6] TTS /api/tts → Khaya TTS (now) → Coqui XTTS (soon) │ ▼ User hears response in their language ``` --- ## Repo Structure ``` ghana-speech-bridge/ ├── apps/ │ └── speech-bridge/ # Next.js API service — deploy to Vercel │ ├── app/api/ │ │ ├── stt/ # Speech → Text │ │ ├── tts/ # Text → Speech (returns audio URL) │ │ ├── translate/ # Text translation │ │ ├── detect/ # Language detection │ │ └── budget/ # API call usage dashboard │ └── lib/ │ ├── providers/ │ │ └── khaya.ts # Khaya AI client │ ├── router.ts …