Children's voices

[woodyard]

My app uses a dynamically created grammar of ten words. Recognition for adult voices works well, but recognition of children’s voices is about 20% lower.

There is an old post regarding keyword spotting here that talks about the use of the kFILLPROB parameter:

https://www.politepix.com/forums/topic/add-key-word-spotting-to-openears-library/

I’m wondering if adjusting this value might resolve the issue?

[Halle Winkler]

Historically, people have had pretty good results by doing some adaptation using kid speech:

https://www.politepix.com/2012/11/19/openears-tip-and-tricks-3-acoustic-model-adaptation/

I don’t know why raising or lowering the filler dictionary probability would affect recognition of speech for a different group of speakers but there’s no harm in trying it.

[Author]

Posts