Picovoice / cheetah

On-device streaming speech-to-text engine powered by deep learning
https://picovoice.ai/
Apache License 2.0
583 stars 67 forks source link

Question on custom keywords/boost #316

Closed tempo-riz closed 5 months ago

tempo-riz commented 6 months ago

Untitled (1)

Here what is the behavior ?

we can see that "a fire ball" and "a-fire-ball" have the same IPA, does it changes something else ? Like the speed of prononciation or other, not an expect on the topic :') I would prefer to use the one with dashes - in all my custom that contains spaces is that okay or a bad practice ?

I tested adding them and boosting them (not at the same time) to see the difference but it seems inconsistent...

laves commented 6 months ago

@tempo-riz, your observations are correct. We treat "a fire ball" exactly the same as "a-fire-ball" - there is no difference in expectations of how it will be pronounced. If you would like one to be differentiated, you can change the custom pronunciation, but other than that it would come down to which word makes the most sense in the context of the surrounding speech.

tempo-riz commented 5 months ago

So just to be sure even if I add/boost "a-fire-ball" the model output will probably be "a fire ball" right ? Because that would make more sense in common speech I guess

laves commented 5 months ago

Yes, that's correct. In this case it would not be beneficial to have both unless you set two different pronunciations.