We need a bit of help here. We are getting ready to release our first machine learning based wakeword tagger and the model for “Hey, Mycroft”, but we need some help to get the accuracy up.
We’ve got a backlog of samples that need to be verified so we can feed them to the algorithm. This takes zero expertise in software or machine learning. We simply need folks to listen to the audio and tag the samples as “Hey, Mycroft”, “Almost Hey, Mycroft or Duplicate”, and “Not Hey, Mycroft”
300 queries takes about 15 minutes. Can you help? Please?
NOTE: This system is up and running again. All systems nominal. Please help! - Feb 21 2018 9:39 AM PST.
NOTE: Our appeal for help needs to be delayed 'cuz we found a bug ( database lockup ). Thanks to everyone who tried! We’ll let you know when it is ready to go again. - Feb 19 2018 5:42 AM PST.
Hey guys, thanks for helping. This is very alpha stuff and the database appears to lock up after a few dozen items are tagged. We’ll look into it and come back here for help again when we’ve found the bug.
Thank you for trying though. I can’t highlight enough how much we appreciate it.
@DarkSide - That is a hugely important point. We are going to launch a website for capturing wakewords independent of the stack and, when we do, we’d appreciate contributions from women and from children over 13 years of age.
Right now the data is heavily skewed toward adult men.
What does the level and coloured bar at the bottom of the screen represent?
It says a valid trigger is “other words then Hey Mycroft”, but I have been marking it as valid if there is any type of background noise as long as I can recognise “hey Mycroft” amongst the noise. Should this be clarified in the instructions?
This page doesn’t seem to work for me… I get a “loading circle” to the right of the yellow button, certain elements don’t load (e.g. the Mycroft logo at the top) and clicking the “play” button does nothing.
I have tried in different browsers (Mozilla FIrefox and Chromium), with/without my VPN and after a restart… Nothing.
actually the server is 200 OKing on that page; I suspect we have a database deadlock again. Not sure how to resolve that one, pinging @Wolfgange who might know more.
Might want to look into this further if it keeps happening.
What did you guys say on Kickstarter? 100,000 words or something like that? That’s a lot of words to be verified by what is likely to be a pretty small number of dedicated users…
Hey all! I have browsed the forum and was not certain where to post this question.
Is it possible to train our own wake words with precise beyond changing the phenome in the web interface? If so is there a procedure for doing so? “Hey Mycroft, Train a new wake word”
I located the following https://github.com/MycroftAI/mycroft-precise-python-experiments and noticed that there is a “train_keyword.py” script. Does this create a model that we can just drop in somewhere. For reference I am using picroft.
Thanks