KWS in Speechbrain

Multilingual Spoken Words | MLCommons has now added a very diverse word dataset to opensource and I am wondering if you guys ever had any thoughts to Keyword?

I have always thought KWS major problem was the lack of datasets of any qty and this has been solved with the above even though maybe they should run again with a emphasis on quality rather than quantity and maybe run of other ASR datasets than just commonvoice.

I am getting good results with KWS but finally having a dataset that I can analyse the phonetics and syllables of the KW to create similar labels so that training and softmax is less of a binary switch as they act as context.
Its quite common to get “Noise”, “KW”, “UNKOWN” labels but the problem with that is “UNKNOWN” as it is unknown and what should go in there becomes a random choice of trying to add anything to stop false positives.
I have created a database on Multilingual Spoken Words | MLCommons with a folderId and a syllable & phone table to do selection.
Now I am not guessing at “Unknown” as an example I will create additional labels of same syllable words with starting phonetics and another same syllable with ending phonetics then all words not in ‘kw’ and what I call the ‘sounds like’ labels and augment up with pitch, tempo, reverb & echo to approx 50k label samples and then training.

Its just a method for any of what you might consider current KWS models from CNN, CRNN, DSCN to att_mh_rnn where I am not saying it has to be the above but I am wondering have you guys ever thought about wrapping Multilingual Spoken Words | MLCommons maybe with tables for phonetic/syllable analysis for various model types?