Will not recognize kid voices


#1

I have Mycroft working, at least well enough for testing. However, two of my kids voices are not recognized no matter how hard they try. We tried careful ennunciation, we tried different wake words. They have standard midwestern US accents. Works fine for me, but only me. Any advice? Some of the wake words had a little trouble with my voice too, so I assume it has to do with accent.


#2

Found one that works for myself and my son:

Hey butter bean
HH EY . B AH T ER . B IY N .


#3

Hey Dan,

Great to hear you found one that is working well. It does depend on which wake word listener you use.

Precise is the default and needs to be trained to utilise new words. This is more work up front but provides better results in the long run. As a larger proportion of our training data seems to come from men, I’ve seen it have a little more trouble with higher pitched voices. This should over time become better. “Hey Mycroft” has been trained much more than the other wake words. Midwestern accents I thought would be the easiest for it to handle, and I’m actually pretty impressed by how well it handles different global accents.

I’m assuming you’ve switched over to PocketSphinx which listens for a set of phonemes eg “HH EY…” so can work be more consistent for words that haven’t been trained but it will never get any better.

There are more details in our docs if you haven’t seen that already.


#4

My take on this is that the children’s voices are not being recognised as Wake Words because their pitch is lower than the range for male voices. Females voices also have a much poorer rate of recognition than male voices as well.

As Gez has pointed out, the key reason for this is that there are fewer children’s voice samples in the dataset that Precise has been trained on.

We need more female and children samples in the dataset.