Delamain

"Delamain online. I trust your day has been satisfactory so far."

[!WARNING] Delamain is a hobby project made for fun. It is not a safety system. Do not interact with any interface, voice or otherwise, in a way that distracts you from driving. The authors accept no liability for accidents, injuries, traffic violations, or any other incidents arising from the use of this software. Use at your own risk. Always prioritise road safety above all else.

A voice companion for the drive, modeled on Delamain from Cyberpunk 2077. It runs in any browser, including the one on your car's Android Auto display, so there's nothing to install. Talk to it and it answers in Delamain's cloned voice. It can search the web, read you the news and the weather, tell you how long a drive will take, speak reminders when they're due, keep notes and long-term memory in your Obsidian vault, and hand navigation off to your phone's map app.

Demo

Screenshots

Features

• Voice personality: formal, measured and dry, exactly as in Cyberpunk 2077. The voice is cloned from a reference clip with F5-TTS.
• Corrupted personality core: switch to a fractured rogue unit under Settings → Personality Core. It has rival sub-personalities, stutters, and a glitched, damaged voice made by post-processing the synthesized audio. Navigation and safety warnings stay intact.
• News bulletins: say "read me the news", "any tech news?" or "local news" and Delamain reads a short spoken bulletin of current headlines. It uses Google News RSS, with SearXNG as a fallback.
• Reminders and timers: "remind me in twenty minutes to call mom" or "at five, remind me to pick up the dry cleaning". Delamain speaks the reminder aloud when it's due. If you're away at that moment, he tells you as soon as you're back. Reminders are kept in the database and survive restarts.
• Weather: current conditions and a three-day forecast for where you are or any town, from Open-Meteo.
• Drive time and distance: "how long to Colorado Springs?" uses OSRM routing at typical road speeds, with an arrival estimate.
• Picks up where you left off: Android Auto drops and reopens the page often. On reconnect Delamain restores the last 30 minutes of conversation, and skips the fresh greeting when you've only been gone a few minutes.
• Knows the hour and your name: greetings fit the time of day, and he uses the name from Settings now and then.
• Web search: traffic, scores, opening hours, prices, places. The LLM calls SearXNG through tool calling.
• Notes and memory: lists, reminders and notes go into your Obsidian vault under Delamain/. Lasting facts about you and the people and places in your life are kept in Delamain/Memory/, one Markdown note per topic, and you can read and edit them in Obsidian.
• Navigation: "take me to the nearest Whole Foods" searches for the place, geocodes it with Nominatim, and opens your map app with the route.
• Voice input: a push-to-talk microphone button using the browser's speech recognition, plus a text chat panel.
• Conversation log: a history panel of everything you and Delamain have said.
• JWT authentication: a single shared password, rate-limited login and 30-day tokens.

Architecture

┌──────────────────────────┐   WebSocket    ┌──────────────────────────────┐
│  Web UI (React + Vite)   │◄──────────────►│  Delamain backend (FastAPI)  │
│  phone / Android Auto /  │  talk, replies │  F5-TTS voice, SQLite log    │
│  any browser             │  audio URLs    └──┬──────────┬────────────┬───┘
└──────────────────────────┘                   │          │            │
                                     OpenAI-compatible  SearXNG /   Obsidian
                                     LLM (tool calling) news RSS    vault

• Backend: FastAPI on Python 3.12, F5-TTS for the voice, SQLite for the conversation log. It works with any OpenAI-compatible chat endpoint (llama.cpp, vLLM, Halogen, …), and the LLM uses the tools in assistant_tools.py.
• Frontend: React + Vite with Tailwind CSS v4 (cyber theme).
• More detail: docs/architecture.md.

Requirements

• Python 3.11+
• A GPU is recommended for F5-TTS (NVIDIA CUDA or AMD ROCm). CPU works, but uncached phrases take noticeably longer to speak.
• Node.js 20+ to build the frontend
• An OpenAI-compatible LLM endpoint with tool calling (llama.cpp server, vLLM, Halogen, …)
• SearXNG (optional; used for web search and as the news fallback)
• An Obsidian vault (optional; used for notes and memory)

Installation

1. Clone and install dependencies

git clone https://github.com/vonhex/delamain.git
cd delamain
python -m venv venv
source venv/bin/activate
pip install -r requirements.txt

2. Download F5-TTS model weights

mkdir -p checkpoints/F5TTS_Base
huggingface-cli download SWivid/F5-TTS \
  F5TTS_Base/model_1200000.pt \
  F5TTS_Base/vocab.txt \
  --local-dir checkpoints

3. Prepare your reference audio

F5-TTS clones the voice from a reference clip. Provide a clean 10-second WAV of the voice you want to clone:

cp your_voice_reference.wav delamain_ref_10s.wav

Edit REF_TEXT in tts_engine.py to match exactly what is spoken in that clip. See docs/f5-tts.md.

4. Configure environment

cp .env.example .env
# Set LLM_URL, LLM_MODEL, SEARXNG_URL, OBSIDIAN_VAULT_PATH and ALLOWED_ORIGINS

5. Set a password

python auth.py set-password <your-password>

6. Start the backend

uvicorn main:app --host 0.0.0.0 --port 8888 --ws-ping-interval 0

7. Start the frontend

cd frontend
npm install
npm run dev

Visit http://localhost:5173 and log in with your password. For production, run npm run build and serve frontend/dist. The full guide is in docs/setup.md.

Docker

The image bundles the web UI, so one container serves everything on port 8888.

cp .env.example .env   # edit .env first
mkdir -p checkpoints data
cp your_voice_reference.wav data/delamain_ref_10s.wav
docker compose up -d
docker compose exec delamain python auth.py set-password <your-password>

Then open http://<server>:8888. Pre-built images: ghcr.io/vonhex/delamain:latest, or a release tag such as ghcr.io/vonhex/delamain:1.4.

F5-TTS model weights and the reference audio are not included in the image.

Environment Variables

API

Details in docs/api.md.

License

PolyForm Noncommercial License 1.0.0. It is free for personal use; commercial use requires a separate agreement.

Acknowledgements

• F5-TTS: voice synthesis
• SearXNG: private web search
• CD Projekt Red: Cyberpunk 2077 and the Delamain character that inspired this project