I have just been looking at Assist - Talking to Home Assistant - Home Assistant and just thinking how great it is to have a inference based control skill supported by the skill project.
My thought was all I now need is a local music library skill for some of us who like tech but not have big data following our lives.
The secret sauce is to be able to select your music ‘Play me some music by The Beatles’…
This has always been one of the tougher assignments in voice recognition due to the wonderful colourful names musicians, bands and music often has against the ML logic.
There has been some talk on Rhasspy 3 Developer Preview to ceate on the fly specific language models based on a specific domain entities.
So with Hass the controls and control methods are chunked into sentences to create a a LM or N-grams.
I am a total noob to Mopidy but due to snapcast and thinking highly of the Dev he seems to like Mopidy , so thought I would ask here.
So apols for all that but has or is anyone writing an inference based skill for mopidy that likely will be hugely more supported than an embedded skill in one of the many Voice Assistants available.
Also just to hint that Language model created from library contents could greatly enhance recognition of what you need to know as both NLU and ASR would benefit being based on the library it will control.
Or just a general SotA ASR like currently OpenAi’s Whisper.