Voice Is Faster Than Calling an API—He Packed an Entire AI Service Stack Into a Desktop Robot
Using M5Stack StackChan plus CoreS3 as the hardware, speech synthesis, VAD, STT, and vision all run mostly on-device, while the brain uses Hermes Agent connected to gpt-6-luna. People are already asking for reproduction steps.
On the second day after getting the M5Stack StackChan, Alex C. (@alexcovo_eth) had the whole service stack up and running. He calls it Hermes-Agent and says 'she deserves to stay.' He had previously used the CoreS3 version of StackChan—white body, with a cartoon girl with black hair running on the screen.
The hardware for this stack is the M5Stack StackChan, with a CoreS3 controller and an ESP32-S3 chip.
The string of services behind it is the real highlight. He uses his Mac as a server and assembled a complete pipeline from local services:
- Voice server: xiaozhi-esp32-server 0.9.6, running locally in Docker
- Speech synthesis: GradiumAI's Phonon, running locally. According to him, this is the fastest local voice he has used—faster than calling ElevenLabs through the API, and free
- End-of-speech detection: Silero VAD, with 0.8 seconds of silence counting as the end of a sentence
- Speech-to-text: local Parakeet
- Vision: gemma4:e4b, running on Ollama
- Brain: NousResearch's Hermes Agent, using the nous-girl agent, connected to gpt-6-luna via sub, with reasoning disabled
He rates the setup as already fairly stable; what's left is fine-tuning. The only thing he's less satisfied with is the camera. He says he originally planned to replace it so it could act as a digital photographer.
In the comments, someone who bought the same robot a week ago and is also using Hermes directly asked him how to reproduce it. Someone else left a comment: 'Pretty good, and also a little scary.'
The whole pipeline is laid out like this—what's missing isn't runnable components, just the will to get hands-on. He left one line at the end of his tweet:
You can just do things.
发布时间: 2026-10-24 19:53