Does Mycroft have trouble hearing female voices?

It’s been maybe 3 or 4 days since I got Picroft working with the AIY Voicekit, and Mycroft hears me pretty well. My partner on the other hand? Not so much. It turns out she has to imitate my lower voice for Mycroft to hear her.

I assume that the reason for this is that Mycroft, like most free and open source projects, tend to be made by males of European descent, and so are most of their users. As such, it stands to reason that most of the voice samples used to develop the STT are from men.

Am I jumping to conclusions? Has anyone else had issues with this?

The wakeword or the rest of the speech?

Also if you’re not contributing to the open dataset with both voices, might want to consider that.


Hi there @teslas_moustache, great question.

The short answer is that yes, we’ve also observed that the Precise Wake Word engine works better for male voice, particularly American accents (as opposed to European accents). For example, I’m an Australian female, with a reasonably deep voice for a female and it doesn’t recognize me as well as my male colleagues.

However, and this is a big caveat, the more samples we have from more diverse voices, accents and genders, the less this will be an issue - because the training data will be more diverse. So yes, as @baconator suggests, please do ensure that you’re opted in to the open dataset.

We’ve toyed with, for instance, tagging samples by gender so that we can reduce what is ingested into the data set for training - for example limiting the number of male samples, but this means de-identifying the data in ways that we’re not comfortable with.



The problem appears to be the wakeword. I looked at some previous threads and found that switching from “Jarvis” back to “Mycroft” is a possible solution. Ironically, she’s the Marvel fan who wanted that. :roll_eyes: of course, it’s hard to tell if it’s just the wakeword or the whole STT because if I’m the only one who can wake it up, I’m the only one who’s likely to give it commands.

So I’ve done that. We’ll see if that works.

I’ll consider opting in, too.

I have this issue to using my wake word “kelsey” K EH L S IY .
I am hoping to train my own wake word at some point. I have an android app on my phone that I am using to gather samples from anyone who will contribute.


Indeed, mycroft doesn’t activates when my partner says “hey mycroft”. It fails 19 of 20 times. English isn’t her main language, so it must affect somehow, but mine isn’t better than hers and mycroft listens to me 100% of times.

I’m NOT using precise, btw. I need to try out precise because Mycroft turn’s on with TV sounds (not even in english) and with random noises, which makes mad my woman xD


It seems to go a little better with Precise, but still far from perfect. She has to deepen her voice in order for it to register.

I have experimented with a custom wake word also using pocketsphinx. The trick I have found is careful modification with it. There are settings which control the sensitivity and such. And, do you have to watch what words you use. I used “hey Ziggy” but I found that was a very common sound while watching television. I ended up changing it to listen Ziggy. And that worked much better. It still takes careful configuration of it. As you can see I am an 80s sci-fi fan. Does anybody know that reference?

Pocketsphinx has far more false positives than I like. Someone [else who has time] should start a precise model builder site and get folks to upload their wake word samples ala mozilla voice and then generate models from that.

Well this just got posted, so that’s a big potential improvement: