Female voice detection, train locally?

If i train my mycroft using the link (below) to help my wife us it, will that help up stream? or will that just be for local wake word and not uploaded to a bigger data set.

Or has anyone tried female voices with the other engines ?

Your precise wakeword will work better with training data from your desired users. If you train it locally it’s saved locally and not used for a larger dataset.

1 Like

That’s a shame, I remember ages ago being able to help train the wake word, by listening to samples and clicking on whether it was the wake word, just mycroft (missing hey) or not either.

Not sure I can convince the wife and 4 year old to say “hey mycroft” 80 plus times.

There’s a trainer skill you can use that should fine-tune the model. 80 might be a bit much, but if you can get say, 10, then do a few things like speed up/slow down <5%, can get a few more copies to start with. Also turn on wake word saving, will help get both good and bad samples.

The problem with the larger dataset that has been reviewed is it’s predominately average north american male voices.

I did think that when listening to the clips, wow they are all men.

I’ll look in to doing the fine tuning. Many thanks

1 Like

Where can one find this skill?

Finally got around to doing this, tho the hit vs miss is about 50%

Is it possible to get my hands on the data that the stock Mycroft if trained on for “hey Mycroft” ?

I suspect the data set is rather large tho

Not yet. There is a newer test precise model for hey mycroft that is supposed to work better for various voices. I think you can search and find a link, i don’t have it handy.

I’ll have a look, i can’t seem to get above 95% when training locally. But i have noticed it has started to her my wife 2/10 times, rather than 0/10.

FYI
The homelab show Hey Mycroft! at the 15:40 they said “the data set is available to the public” with regards to training the wake word.

https://thehomelab.show/2021/09/10/the-homelab-show-ep-24-hey-mycroft/