Goshi - Voice AI for Reachy Mini

🎙️

Wake Word

Say "Sushi" to activate the robot

🤖

Smart AI

Natural conversations powered by LLMs

🌍

Bilingual

Speaks French & English

⚡

Low Latency

<800ms response time

🔒

Local AI

Private, on-premise processing

🤸

Expressive

Head & antenna animations

🚀 Installation

Install on Reachy Mini

Via the app manager or command line:

pip install git+https://huggingface.co/spaces/pnocera/goshi

Start the Goshi Server

On your workstation (requires NVIDIA GPU):

./goshi-server --config config.toml

Configure & Talk!

Update config.yaml with server IP, then say "Sushi"!

🏗️ Architecture

  Reachy Mini (Raspberry Pi 4)          Workstation (RTX GPU)
  ┌────────────────────────┐           ┌─────────────────────────┐
  │       Goshi Client     │◄── SRT ──►│     Goshi Server        │
  │  ┌──────────────────┐  │           │  ┌───────────────────┐  │
  │  │ Wake Word (KWS)  │  │           │  │ VAD (Silero)      │  │
  │  │ Audio Capture    │──┼───────────┼─►│ STT (Whisper)     │  │
  │  │ TTS Playback     │◄─┼───────────┼──│ LLM (Ollama)      │  │
  │  │ Video Stream     │  │           │  │ TTS (Piper)       │  │
  │  └──────────────────┘  │           │  └───────────────────┘  │
  └────────────────────────┘           └─────────────────────────┘