Gemma 4 12B QAT + faster Qwen 3 TTS fits perfectly fine in 12GB VRAM and can be used to create a fully local voice assistant with … reconizeable traits ;)
The project is fully open source and published here: https://git.sync.wtf/starfleet
I see Youtubers trying to do similar things every now and then but for some reason their latencies always seem a lot higher than I would expect so if the one I’ve conjured up (based on Huggingface’s Speech-to-Speech project initially) is special somehow then maybe it’s of interest to others.
Nice project!
Btw, just to let you know, in case your “computer” model is still to be trained… (as you write in the Readme) There’s a bunch of people who make microwakeword models for ESPHome and put them on GitHub:
Thanks, I need to update the README :D Implemented full training support in the project yesterday so I could use “computer” instead of “hi ESP” :D
Awesome. I’m still looking for some open-source solution which does pretty much what you do, plus a graphical avatar. I don’t have an exact use-case but I imagine it’d be nice as an AI Vtuber, maybe some virtual creature (I got a few unused displays for ESP32s in my drawer) or some sort of desktop assistant / secretary.
What I currently do is use ESPHome and Home Assistant. I got a few satellites around the house and they can do speech commands like change the light, turn on the TV or set a timer or inform me about the weather conditions… Unfortunately I’m lacking a GPU in my server so all of that is a bit more dumbed down to keep the latency down. So no tool calling or LLMs or any fancy stuff. Requires me to say fixed commands, but in turn makes it incredibly fast, even on a CPU. It’ll do for smarthome application, but I suppose that’s a different scope from what you’re programming.
This solution is fully open source - and the components used for SST/LLM/TTS are freely exchangeable. Latency goes down if one has the ability to use a separate SST like Parakeet instead of reusing the LLM for it as I do in this video.
My scope is “what me and family wants” - Home-Assistant is one of the tools the LLM lists (pool pump sensor comes from there). The server takes an mcp.json so any tool that has an MCP connector can be used.
Haven’t thought a graphical avatar, but I am a Red Dwarf fan and I’ve already thought about the ability to have pre-configured “themes” for the system (Star Trek, HAL, Holly etc) - although not something I can host and publish due to … waves hands … IP rights.
Ooh seems great !
I am nowhere near my compter (hah !) right now so I can’t test, but last time I checked, qwen3 TTS was NVIDIA (well CUDA) only
Will my full AMD rig on linux be able to run it ?
Ooh, we’re getting close to my dream scenario now! Just one bit left, though. Which model do I install to make my voice sound like Captain Picard’s?
I assume you meant your and not the computer’s, and I agree - would also like to sound like him sometimes! :D
For those interested in the voice cloning part of the project though, it’s done “live” at server startup through two config options:
--qwen3_tts_ref_audio ./computer.wav --qwen3_tts_ref_text "Darmak is the name of a seventh dynasty emperor on condon four A myth of a historical hunter on Chantil three A colony on Melindi seven"Selecting the voice of the computer thus only needs a clear .wav of them speaking and a transcription.
