Wake Word Trainer Skill


#1

Wake Word Trainer Skill

In order to train a new wakeword you have to have read very far in mycroft. in the endefekt but deas wakeword is only a few times spoken and bad detection sorted out. You could not do that through speech. I have thought about it but have too little idea to do it myself.
conversation:

  1. Hey mackroft i want to call you christopher.
  2. No problem. Please say after each of the ten piep christopher.
  3. “Piep” christopher “piep”…
  4. please wait ten minutes while I compute what wakeword.
    wakeword improves:
  5. you do not understand me right yet.
  6. I will show you some wakewords please tell me yes or no.
  7. “Wakeword” no.
  8. “wakeword” Stop.
  9. please wait ten minutes while I compute what wakeword.

that sounds realizable to me. for example, the audio clip from mimic-recording-studio shows very well cut audiofiles. but there is a lot of work like changing mycroft.conf and compiling pricice.


#2

I agree, you can build a skill that works as frontend for Mycroft Precise.
I didn’t train my own Precise wakeword yet, but when I read el-tocino tips on precise wake-word I think it will be more like hundreds or thousands recordings than only ten…


#3

I’d guess fifty is the start of the sweet spot. Not-wake-words seem to be as if not more important, but there’s ways to generate or source those from existing data as well. For best effect specific nww’s would be ideal, of course, and sourced from target audience.


#4

Also helpful if you’re into experimenting is saving your wake words. The tips link above has an uploader utility that can be run on a local computer to your mycroft instance (or even on the mycroft instance) and some config that can be used to do wakeword uploading. Turning that on will get you all your wakewords it hears, which can expand your dataset quite well, particularly if you’re getting a lot of false positives.


#5

I started with the first attempts. Thanks for the note I will have to increase the number. which leads to the fact that one must be able to break off the reading process. only when a certain number is reached does the calculation begin. I have to deal more closely with the tool.
if someone wants to help me he can get in touch.


#6

I think the skill is a good idea, but the volume of data without supplemental help is the bigger hurdle. You can make a precise model that beats pocketsphinx without much data, but to make a more reliable model is where it gets more involved.