search

MycroftAI / mycroft-precise

A lightweight, simple-to-use, RNN wake word listener
Apache License 2.0
848 stars 226 forks source link

Trained model doesn't trigger #58

Closed torntrousers closed 5 years ago

torntrousers commented 5 years ago

I've trained a model with the phrase 'lamp on' following the guide here but then using the result I can't get any triggers when running precise-listen and saying the phrase 'lamp on'. Could I have done something wrong somewhere? How can I debug this? The generated model is zipped up here, can you see anything from that?

el-tocino commented 5 years ago

in your "test" directory, you should have a "wake-word" as well as "not-wake-word" directories, with each populated by the relevant files.

You should add a significant amount of not-wake-word data, particularly words/phrases that start similarly, and ones that contain similar phonemes. (camp on, crampon, lantern, amped up, etc). A mention over on chat by matthew indicated a 1:5 wake:fake word ratio was their target. More data in general isn't a bad idea for the whole thing, either.

When you train, run precise-train (options) && precise-test (model) to get an idea how it performed on your data.

torntrousers commented 5 years ago

Ah yes, having the folders lamp-on/test/wake-word/ does fix it. Thanks!