JMS1717/ming-voice
LOCAL SPEECH-TO-SPEECH ASSISTANT ON GPU
Fully local voice assistant with ~0.23s latency from end of phrase to answer. Runs on AMD/Nvidia GPU under Windows and Linux without sending audio to cloud.
// KEY FEATURES
- ~0.23s latency with Smart Turn v3 for turn-end detection
- Home Assistant, Jellyfin, timers, lists, and memory integration
- 7 voices + cloning from 5-15s sample, OpenAI-compatible API
- Runs always-on, phone access via HTTPS, Wyoming protocol
// DEEP DIVE: HOW IT WORKS
Ming Voice wraps the inclusionAI/Ming-3.0-audio-tiny model into a local voice assistant running on the user's GPU. Audio is processed entirely on-device: the system listens to the microphone, detects end-of-turn via Smart Turn v3, generates a response and speaks it back in ~0.23 seconds. Supports interruption, pause-and-continue, and persists context across restarts.
Inside is the Ming-3.0-audio-tiny speech-to-speech model, optimized for AMD (ROCm) and Nvidia (CUDA/Triton). It stands out with built-in tools: web search, weather, calculator, timers, reminders, shopping lists, and long-term memory. Deep Home Assistant integration enables control of lights, climate, locks, media, vacuum, and scenes, including scheduled and grouped commands. There is a Jarvis mode with a distinct persona and voice, voice cloning from a short sample, and an OpenAI-compatible API for TTS, STT, and audio chat.
Requires AMD GPU (ROCm) or Nvidia (experimental), Windows or Linux. Models download on first run; only text queries leave for web search and weather. Audio never leaves the machine if tools are disabled. No browser UI — runs as a background process with HTTPS phone access and Wyoming integration for Home Assistant Assist.
// USE CASES
Best for: For local AI and smart home enthusiasts on AMD/Nvidia GPU
- Control smart home by voice: lights, climate, locks, media and Home Assistant scenes
- Maintain shopping lists, timers and notes offline with context persistence
- Run local voice assistant on PC and use it from phone via HTTPS
New repositories every 30 minutes
REDDYX AI scans GitHub 24/7 and ships the best AI/ML/Web3 projects to Telegram.
Join on Telegram