So I have been playing with librespeaker getting it to work for snips with minimal external dependencies and I have noticed a problem (or at least something weird when I use it) with multiple Snowboy hotword models.
When I provide multiple models to the Snowboy1bDoaKwsNode::Create call and then provide matching multiple sensitivities I have to provide one extra sensitivity for it to work.
An example:
I provide the following for my hotword model string:
</s>Olga_Chris.pmdl,snowboy.umdl,computer.umdl,jarvis.umdl,alexa_02092017.umdl,OK_Google.pmdl<e>
and the following sensitivity string:
</s>0.4,0.4,0.4,0.4,0.4,0.4<e>
So 6 models and 6 sensitivities in total.
I get the following error when running:
Strangely its saying I have 1 extra universal hotword then I have provided. If I add an extra ,0.4 to my sensitivity string then it no longer gets unhappy.
I also see a strange hotwork activated index when I trigger the hotword. The index doesn’t seem to match up with what model I expect to be triggered.
After doing some more digging it looks as if its possible for snowboy universal models to actually contain multiple hotwords.
A quote from the snowboy-detect.h include file: </s>// CAVEAT: a personal model only contain one hotword, but an universal model
// may contain multiple hotwords. It is your responsibility to figure
// out the index of the hotword. For example, if your model string is
// "foo.pmdl,bar.umdl", where foo.pmdl contains hotword x, bar.umdl
// has two hotwords y and z, the indices of different hotwords are as
// follows:
// x 1
// y 2
// z 3<e>
So after process of elimination it turns out that the jarvis.umdl model that I was including actually contained two hotwords (not that thats documented any where), and that was skewing the detected hotword index. It was also the cause of the error message saying I had not provided enough sensitivity values.
I think this issue can be considered solved. But I’ll leave this here so people can reference it hopefully if they run into the same thing.
One note I would add is that I think you guys (seeed) should add similar verb-age to the librespeaker documentation so that its a know thing and people don’t get tripped up on it when using multiple models (especially universal ones that may contain multiple hotwords).
