JemmaUZH/jiema-wolof-assistant

# Wolof Audio Assistant A small browser demo for a Wolof-first assistant for Senegalese users who may prefer Wolof over French. Jiema is being built toward a local, low-memory voice stack. ```text Wolof voice or transcript -> Wolof ASR -> small local LLM -> simple Wolof answer -> optional Wolof TTS ``` The app does not require a domain selector. It accepts Wolof, French, and mixed Wolof-French input, then asks the model to infer whether the request is about agriculture, transport, health, administration, education, finance, or general daily help. See PROPOSAL.md for the product/demo proposal. ## Run ```bash npm install OPENAI_API_KEY=your_key npm run dev ``` Then open: ```text localhost ``` Without `OPENAI_API_KEY`, the original Whisper path is disabled. Text answers can still run through the local backend paths described below. ## Local Pipeline The current local pipeline server is experimental and runs on your machine: ```bash python3 server/jiema_pipeline_server.py ``` Default local models: ```text ASR: sokho2/mms-300m-wolof LLM: google/gemma-4-E2B-it-qat-mobile-transformers TTS: facebook/mms-tts-wlx ``` Then point the Node app at it: ```bash JIEMA_PIPELINE_URL=127.0.0.1 npm run dev ``` Useful local endpoints: ```text GET /api/health POST /api/warmup/all POST /api/answer POST /api/transcribe POST /api/synthesize ``` Mac prototype note: `google/gemma-4-E2B-it-qat-mobile-transformers` loads through Transformers/MPS on the tested Mac, but text generation currently returns ` `. The model remains the mobile/QAT target, while the Mac path is mainly useful for memory and integration experiments until the runtime issue is fixed. ## Legacy Local Chat Endpoint Jiema also supports an Ollama-style local chat endpoint for earlier Sunflower/Gemma experiments. The backend calls an Ollama-style `/api/chat` endpoint: ```bash LOCAL_CHAT_MODEL=Sunflower-Gemma4-E2B \ LOCAL_CHAT_URL=localhost \ OPENAI_API_KEY=your …