My Utopia (Some dreams )

Hi there.

First of all , thanks you for your amazing Work !

Like a lot of people there , i was using Snips before sonos arrived …

I try rhasspy with a matrix voice + some raspberry , it’s working nice !

I set French profile , i try Kaldi / PocketSphinx with and without mix_weigth or open_transcription.
My backend Node-red Handle intent even for home-assistant.

My Utopia is to control everything at Home from a self-hosted assistant (without external stuff… at least if i can ^^)
Still dreaming !

So control :

  • EASY : Home assistant
  • EASY : Logitech harmony (throught HA)
  • EASY : Weather (throught HA)
  • EASY : internal Temperature (throught HA)
  • EASY : my xbox stuff (throught HA)
  • EASY : Other stuff (throught HA)
  • EASY : PiHole (via node-red)
  • EASY : Synology (via node-red)
  • EASY : (via node-red)
  • MEDIUM : Logitech media Server Squeezzbox (Max2Play)
    Maybe via HA or directly by API in node-red
    Need open_transcription or Big Slots auto-generated
  • MEDIUM : Spotify
    directly by API in node-red
    Need open_transcription or Big Slots auto-generated
  • MEDIUM : Kodi
    directly by API in node-red
    Need open_transcription or Big Slots auto-generated
  • HARD : Streaming Stuff (Netflix / MyCanal)
  • HARD and WAF : Recipe for me and my wife via external display …
  • HARD and WAF : Shooping List (Via HA or Grocy …)

Here is my thought (after this long intro ^^ :sleeping: …)

I’m looking for real open transcription to handle Spotify Intent like :

’ Listen some Muse in Spotify’

First try with kaldi or sphinx without open_transcription don’t work cause the Artist / Album / Playlist slots doesn’t exist and i think it’s really hard to generate !

Second Try , open transcription with kaldi :
Better !
I can say some artist in French , and it’s working … but now , Spotify words doesn’t exist .
So it’s not recognise … FUUUU

Third Try , open transcription with pocketsphinx:
i got some pocketSphinxDecoder error …

Last : mix_weight
Long time to STT and result is not what i’m looking for …

Other question in mind :

Do you have a way to handle multi-turn intent or conversation ?

I may used some work around for complex and unatended intent like

Ask first general Stuff like ‘listen artist in spotify’
Rhasspy decode it pass to node-red who ask me which artist want to listen (with another SST engine ??? )

I read a lot of topic fro that forum. I would like to contribute as I can . Maybe share some Slots .
Share some custom Words list …

Thanks !!

3 Likes

Had a look at this ? : https://rhasspy.readthedocs.io/en/latest/training/#slots-lists

Yes ! Of course ,i’ve read everything from the documentation .

It’s a option but when you want to play song from Spotify , the slot list will be aweful and giant…

So the idea : have some unknow word will be catch and handle by intent like artist , album , song …

I’m trying rasa now to handle unknow word .

I’ve set for now Google SST to have the real word speech !

I may Switch to kaldi After that .

Yeahhhhhhhh !

I got it working… Not so good but …

So I use Google stt to have a perfect and non privacy transcription ( handle by node red via API call from rhasspy)

Then rasa nlu with custom config also call from rhasspy via node-red ( custom parser etc…)

And intent handle to Spotify via (again ) node red !

I can search tracks by name , artist send to any Spotify connected device :wink:

I’m soooooo happy

Need to train a lot more but rasa x as a special conversation from rhasspy and I can as long as I used it , correct every sentence and find some way to handle it !

Poyaaaaa

1 Like

4 Likes

This looks awesome! Would you care to share your setup ?

1 Like