#Trouble connecting Voice Assistant Preview to my local setup.

1 messages · Page 1 of 1 (latest)

dull whale
#

Let me describe my setup: raspberry pi 4 with HA, Mac Mini M4 Pro with Ollama, Piper and Whisper running. Assist works in text mode, i.e. I can ask it to turn the light on, and it does, so that's nice. I can't test microphone or speaker because not using HTTPS for HA.
I bought and installed the little preview box, connected it, and sometimes it works, but it doen't talk back properly. It seems to speak english using a dutch voice.
I ran a network traced on port 10200 (piper) and it seems to be emitting JSON intermixed with audio. Here's a piece of the piper.err file:

DEBUG:wyoming_piper.handler:Synthesize(text='Since the user made a typo, but the context suggests they meant lights, the response should be straightforward.\n</think>', voice=SynthesizeVoice(name='nl_NL-mls-medium', language=None, speaker=None), context=None)
DEBUG:wyoming_piper.handler:synthesize: raw_text=Since the user made a typo, but the context suggests they meant lights, the response should be straightforward.
</think>, text='Since the user made a typo, but the context suggests they meant lights, the response should be straightforward. </think>.'
DEBUG:wyoming_piper.handler:Synthesize(text='Verlichting uitgezet', voice=SynthesizeVoice(name='nl_NL-mls-medium', language=None, speaker=None), context=None)
DEBUG:wyoming_piper.handler:synthesize: raw_text=Verlichting uitgezet, text='Verlichting uitgezet.'
DEBUG:wyoming_piper.handler:Text stream stopped

Any pointers?

crude pendant
dull whale
#

qwen3:4b I think, new to ollama

crude pendant
#

the stock 4b seems to have an issue with thinking getting forced on and into responses. instruct is better suited for voice assistant anyway doesnt have the think issue.

#

In fact the default model home assistant selects is changing over to it. THIS got merged yesterday and will be included in next version.

dull whale
#

Oh, this works better indeed, it now actually said something back (that it couldn't understand me).

#

I'm Dutch, it's bad in understanding my Dutch. Can I improve that? I have plenty of power available.

crude pendant
dull whale
#

This is my startup script:

exec python -m wyoming_faster_whisper \
  --model small \
  --language nl \
  --uri tcp://0.0.0.0:10300 \
  --data-dir /Users/ha/whisper-data \
  --download-dir /Users/ha/whisper-data \
  --device cpu \
  --compute-type float32 \
  --beam-size 5
crude pendant
dull whale
#

I will. I could not get the GPU support to work on my M4, any pointers?

crude pendant
#

i dont know about mac compatibility

#

you could try and find a middle ground too

#

i think the models are:
tiny/base/small/medium/large/turbo

#

turbo is large but optimised basically.

dull whale
#

OK, thanks, I'll play with it.