AetherChat is a lightweight voice-chat API that converts user speech into text using VOSK, processes the request through an Ollama-hosted LLM, and responds using Kokoro TTS. A FastAPI backend (served via Uvicorn) powers the API, and the repository includes a browser-ready web UI.

• Speech-to-Text (STT) via VOSK
• Locally hosted LLM through Ollama
• Text-to-Speech (TTS) synthesis using Kokoro
• FastAPI backend with automatic streaming responses
• Caddy reverse proxy with optional HTTPS
• Fully containerised with Docker Compose
• Customisable voices and model selection

docker-compose.yml defines two core services:

• Builds the Python backend
• Mounts models/vosk and kokoro
• Runs FastAPI on port 8000
• Exposes it externally as 8888 (avoiding conflicts)

• Fronts the API with HTTP/HTTPS

• Serves static UI files

• Proxies inbound traffic to the backend

• Generates internal TLS certificates automatically

• Listens on ports:

  • 8080 (HTTP)
  • 8443 (HTTPS)

Shared settings like OLLAMA_HOST, LLM_MODEL, and VOSK_MODEL_PATH are passed to voicechat. The DOMAIN environment variable controls Caddy’s routing (default: voicechat.local).

Place the required models in the following folders:

./models/vosk/
./kokoro/

(Include Kokoro binaries if you want TTS enabled.)

If you're using the default internal domain:

voicechat.local   →   127.0.0.1

docker compose up --build

Caddy will automatically generate and trust a local certificate (via tls internal). Open:

• https://voicechat.local:8443 – secure UI/API
• http://voicechat.local:8888 – plain HTTP

When using tls internal, Caddy acts as a local CA. You must install its certificate on clients before Safari/iOS/macOS will allow microphone access.

Extract it with:

docker cp voice-chat-caddy:/data/caddy/pki/authorities/local/root.crt ./caddy-root.crt

1. Double-click the certificate
2. Add to “System” or “Login” keychain
3. Set Always Trust

1. Airdrop/email the .crt
2. Tap to install
3. Enable trust under Settings → General → About → Certificate Trust Settings

Once trusted, microphone access will work without warnings.

The default Kokoro pack exposes several selectable voices:

Add more voices by extending the Kokoro .bin file and registering them in app.py.