all 6 comments

sorted by: hot top controversial new old
[–] 12 points 4 months ago (1 child)

Is this the last part I need to actually now bin the Google Home devices?

I use them to switch lights, listen to music and change the temperature.

With this, could I use that as the voice commands into home assistant for example?

  • source
  • hideshow 1 child comment
  • Man i would like to be this technical or have the time to tinker around

  • source
  • [–] 1 point 4 months ago* (1 child)

    I'm trying to set up an always transcribing speech to text note taking app on a pixel 6. Does anyone know a good place to start on this?

    It's wifi enabled but I guess by "sim card doesn't work'" they meant carrier locked. So I can't get graphene on it. But I figure if the wifi is off it won't phone home.

  • source
  • hideshow 1 child comment
  • [–] 1 point 4 months ago

    Neat. But this is all doable on home assistant currently. I was hoping the OpenAi part had been solved but sadly it has not.

  • source
  • 3 comments in c/selfhosted@lemmy.world view that post
    [–] 13 points 4 months ago (2 children)

    Im sure this works but from my recent experience you need your STT on a machine more powerful than a PI atm. Tiny models are just not accurate enough and the bigger ones need more than the PI has to give any sort of decent response time. Compared to where this was two years ago I look forward to where it is in two more.

    One of the largest improvements imo has been microwakeword and the ability to run the hotword detection “on device” but I believe it only runs on ESP32 devices so not an option if want everything on a pi.

    I spent a little bit of time getting a fully local voice pipeline setup in home assistant last month and I’d say it is near perfect(after adding a few additional community integrations) with the exception of the microphones on current hardware. I look forward to the next HA voice device from Nabu Casa.

  • source
  • hideshow 2 child comments
  • [–] [S] 6 points 4 months ago (1 child)

    OpenWakeWord (the solution I also use in my Platypush setup) is actually very good at hotword detection, especially when paired with libspeex for noise suppression.

    The STT models from Vosk are actually quite good too. The 180 MB English model is light enough to run on a RPi and it's actually quite decent with clean accents. They also have a 1.8 GB model that is very accurate and that actually can still run fine on a RPi5 8 GB.

  • source
  • parent
  • hideshow 1 child comment
  • [–] 4 points 4 months ago

    Don’t misunderstand. Openwakeword works great. I just think it’s awesome that the hotword detection can run so well directly on low powered devices now.

    About STT I’ll agree that they can run well, I just found the experience a lot better on a heavier model running on a beefier machine. They do well in a silent space but struggle when you add background noise.

  • source
  • parent