A homeassistant Whisper / Piper docker-compose designed for use with strix halo

Framework Desktop
1

I had trouble finding ROCm (especially latest ROCm, i.e. 7+) setups of whisper and/or piper. So I put together a little setup in this repository. Hope it helps others looking to run the HA voice stack locally on their new Framework Desktop (Mainboards)!

4 Likes
2

Just added support for Qwen3-TTS wrapped in Wyoming as well as Pocket TTS wrapped in wyoming!

Unfortunately qwen3-tts is not generating fast enough to be considered for voice assistant use. It does work, however, specifically with the Qwen3-TTS-12Hz-1.7B-VoiceDesign model which allows designing the voice via a prompt at startup. An audio snippet for a ~10 word sentence took anywhere from 30s-1min to generate.

I think this is mostly down to missing flash-attn in the container. I wasn’t able to get this to work with ROCm 7.1.1 nor 7.2.

Pocket TTS on the other hand works a treat! It is a good bit better than piper in terms of generating natural sounding speech. You can play around with the available voices in their demo page here.

3

I made kokoro-fastapi work on Strix Halo if your’re interested in TTS: Add permutation that support running on ROCm gfx1151 devices such as Strix Halo by projects-land · Pull Request #431 · remsky/Kokoro-FastAPI · GitHub

(At the moment when I made it work ROCm 7.2 hadn’t been released yet so it uses nightly. Now it should work with stable 7.2 ROCm.)

4

Nice one! Any idea if these Strix Halo boxes can generate audio at ~real-time rates via Kokoro?

5

It is fast even on CPU. ON GPU it is very fast. Give it a go.

6

I actually had a Kokoro-FastAPI ROCm fork running already, but it turns out it was using the CPU still :joy:.

I got your patch running with some minor tweaks (commented in the PR) and wow - its definitely useable for voice assistant real-time use-cases now! :heart:

7

What would you recommend (in terms of where to start the learning process) to be able to implement this? I am a quick learner but relatively new to running HA voice stack locally…

thanks!

8

Hmm best to probably learn the basics of docker and Linux sys admin stuff first. But if you’re reasonably well versed in those - just dive in! Try to run some models. Ollama and llama.cpp are good starting points