jannislehmann / HomeAssistant

4 stars 0 forks source link

Build a self-hosted voice assistant #16

Closed jannislehmann closed 4 years ago

jannislehmann commented 4 years ago

I plan on ditching all the Amazon Echo devices and replace them with Raspberry Pis, which use Rhasspy.

The first test is working great so far. I am hosting a master Rhasspy server on my main server and use a raspberry pi zero as a satellite with a PlayStation 3 Camera as a microphone. Everything works so far. There is quite a delay, after using the hotword.

ToDo:


Motivation

The main motivation was, that I wanted one multi-room-audio solution. Currently I have some Echo Dots, one Sonos One, and one dumb soundbar. Therefore, the Echo Dots do have Multi Room Audio and I could use an Echo Dot for the soundbar, but naah. I can't stream via AirPlay or use Mopidy to play YouTube Audio on those. They are simply locked in. Alexa is great though, but I don't need all those weird features. I only want to turn lights on and off and tell it to play music. The music part will be interesting as the PlayStation 3 Eye Camera microphones have no active echo cancellation.

jannislehmann commented 4 years ago

Hass.io addon

Works great. The only thing that does not work is the internal add-on communication. That might be, because rhasspy launches containers itself. Therefore, the new containers are not on the internal hassio network.

jannislehmann commented 4 years ago

Intent Recognition

I can't get it to work that the master server sends the request to HA. It does register the intent and forwards it to Node-RED, but not to HA. There was a thread about that issue https://community.rhasspy.org/t/home-assistant-rhasspy-server-satellite-issue/1195 but I couldn't seem to fix it. I will leave it for now and migrate the native intents to Node-RED at a later point.

jannislehmann commented 4 years ago

The satellite seems to die after some minutes and only works after a restart again. This has to be resolved.

jannislehmann commented 4 years ago

The PoC was working great except from the hotword detection ratio. However, I no longer see the necessity in a voice assistant at all. Might re-open this, when a speaker solution has been found.