Kaldi Error When Trying to Use Massive Slot List

Help
1

I have got an enormous slot list of Spotify artists (~80K rows) of the form

(Nate Dogg):(spotify:artist:1Oa0bMld0A3u5OTYfMzp5h)
(Ron Browz):(spotify:artist:1iqf0x89oxmmAFv9w9hiJM)
(The Streets):(spotify:artist:4GvOygVQquMaPm8oAc0vXi)

I can successfully save my slot. However when I try to save the intent using the slot I get the following error:

TrainingFailedException: Command '['fstcompile', '--isymbols=/profiles/en/kaldi/model/data/lang/words.txt', '--osymbols=/profiles/en/kaldi/model/data/lang/words.txt', '--keep_isymbols=false', '--keep_osymbols=false', '/profiles/en/kaldi/language_model.txt', '/profiles/en/kaldi/model/data/lang/G.fst']' returned non-zero exit status 1.

As far as I can tell I have fully cleaned the slot text prior to loading it to Rhassy, including removing null rows and special characters.

I tried to Google the error but couldn’t find anything that helped.

Does anyone know what this error is telling me? Do I possibly still have a special character hidden somewhere in my slot list?
Is it even possible to use such a large slot list? If not then could I split it into say 20 smaller slot lists and have multiple intents for each variation? Or even use a slot of other slots?

Thanks

2

I’m pretty sure that a gazetteer slot with that many items will never work (reliably or at all). Splitting it won’t help.

I can only advise you to greatly reduce the number of values in this slot.

A few hundreds carefully reviewed might be ok.

PS: You’ll probably need to work on the custom pronunciations of these artist names also.

3

Hi @fastjack,
I had a slot list of about 1200 artists and it performed fine on a 4gb pi 4.
The main problem I had was the list I was using contained a lot of rubbish and without fail if a band name wasn’t quite recognised correctly then some awful music would play! I’ve cut my artist list down to about 300 good artists so that should solve that issue (or at least only play something decent when an artist name is incorrectly recognised.)

Re pronunciation, it has been pretty good in the most part. The only thing it has really struggled with are band names that are initials eg NWA, AFI, but otherwise it’s pretty decent out of the box :slight_smile:

Out of curiosity, are you aware of a better way to interact with Spotify via Rhasspy? I would love to be able to search Spotify and play any artist / album / song. Currently investigating if it may be possible using Spotipy but not far enough into it yet to know if it will be possible. Also using LMS / Picoreplayer with the Spotty plugin and the material skin if that offers any possibilities?

Thanks