Currently I have Snips running in my house with one RPi 4 running the master and several Pi zero satellites. Sadly Sonos took over Snips and now I’m looking for an alternative that allows me to reuse most of my hardware.
I want to control home assistant thorugh this and this project looks promising. Was directed here from somebody in the Home Assistant community.
My question is where do I find some kind of documentation or something? Seems that this website only has this forum?
For text to speech picoTTS sound decent and works without problems.
Good Wakeword detection was the hardest part for me. I am using precise at the moment as it works best for custom wakewords, but that does not seem like an option for you as it does not run well on a pizero(if at all ).
Snowboy with the wakeword snowboy is usable for german from my experience if you set the sensitivity relatively high. Wakeword Choice for German
I am also using Kaldi and PicoTTS (it’s OK, but compared to Google or Amazon it is still really bad ).
Wake word is an issue. I am currently using porcupine with one of the universal models. Problem: I don’t really like the wake word choices. All feel weird to say for native german speaker and not easy to pronounce without some German accent. I am currently using “pico voice” as wake word. It recognizes me most of the time and very few false activates.
@moqart
I am thinking about making my only custom wake word with precise. How did you do it (source of the training data?)
Have you tried MaryTTS? I don’t know the quality of the German voices, but I’m using the voice dfki-prudence-hsmm for English and it’s really enjoyable.
@DanielW If you want more information on training a custom wakeword model with mycroft precise take a look at this Mycroft Precise Model .
I followed the instructions from mycroft for training your own wakeword model.
I wrote down some of my experiences about training a custom wake word here in this post Mycroft Precise Model Problem (computer-en.pb) especially with training a noise resistent model.
I just tried the three German voices but there is no huge difference in quality compared to PicoTTS for me. But it uses more RAM and CPU. (it is OK on a Pi 4. But longer texts will slow down Rhasspy)
I use the PSeye cam on the big Raspi which is working great. Later I want to try my zeros with a two mic platine attached (may check for the name later) which was working great with snips before. I think those satellites have an LED I might start.
I found the respeakers just to expensive when creating multiple devices. Mist are even placed where they are invisible