Trained wake word model only works when speaking slowly [openwakeword]

[sour swan]

Hi! I followed this guide on creating a wake word model for the word Calipso. The only issue is that it works only when speaking it very slowly. I have included a video of testing, as well as some of the parameters.
Is there anyone more experienced that could help me with this ?

[mortal trellis]

For those wake words to work better, TTS samples isn't enough, unfortunately. That guide is more for educational purposes, not for real usage...

[sour swan]

Ah...so no way of getting any usable result without providing real training data

Around how many recordings of people saying Calipso would I need to train a good model ?

[mortal trellis]

IDK, Okay Nabu used 10000 samples so far.

[sour swan]

holy shi...

yea...no way i'm making my own thing

[mortal trellis]

Well if you gather samples from you and your family (couple hundreds to a thousand), it definitely will help! 🙂

[sharp bay]

Think the last model that was trained had 20k good samples after filtering the 30k total samples available

[mortal trellis]

Oh, right, that one 1223?

[sharp bay]

I think, was the last one Kevin did i think

Yeah 1223

Think 1226 is on the way from that conversation though

[mortal trellis]

Yup

[chrome torrent]

20k+ samples are needed for a robust universal model, but as formatBCE said, your own samples are far more useful and not needed in such large numbers @sour swan

[sour swan]

gonna ask the people most likely to use the assistant for around 20 each, and i'll submit 100 of my own...hope for the best

is there any training script for my own audio ?

[sour swan]

all i found online is the TTS method lol