Voice commands for iOS app

[Alexey]

Hi,

I think this is a probably very common usage for OpenEars as dictionary size is very limited.

I did all the code according tutorial and I found that app tries to recognise my speech using only words from the dictionary. So if I have 3 words in dictionary anything I say will be chosen from this dictionary.
Is it right?
If yes – I completely didn’t get what supposed recognitionScore in pocketsphinxDidReceiveHypothesis:recognitionScore:utteranceID: mean. It just assign random values. If I say actual word from the dictionary it can be -4000, or -2000 or 0. If I say word not in dictionary hypothesis is still word from the dictionary and recognitionScore can be even closer to zero than actual word…

Am I right that for my case (I need to recognise 3-5 commands, single words) I must include Rejecto?
Is there any way to improve recognition of reject (at least make it stricter, I’m ok if it won’t recognise correct word but I don’t need false triggering).

Cheers!

[Halle Winkler]

Welcome Alexey,

This is the usage case that Rejecto was designed for – the scores in general are not very informative with a language model that small, unfortunately.

[Alexey]

Thanks, Halle.

That what I understood.
Am I right that there is no way to make Rejecto more strict? Or “teach” it somehow?

[Halle Winkler]

You can raise its weighting versus the rest of the language model – have you given that a try?

[Alexey]

Nope. I didn’t know how to do that. I’ll dig the documentation. I just noticed that with Rejecto all scores became zero, so I thought that all guesses are equally weighted now.

[Halle Winkler]

There’s a weight argument in the main Rejecto method for language model generation, so I’d check it out and see if it accomplishes your goal. If you’re seeing questionable accuracy please make sure you aren’t testing on the Simulator.

The scores really can’t be relied on for any logical purpose in general app code – if I hadn’t forwarded them to the API in the early versions due to a wish to replicate the research-related functions of Pocketsphinx I would probably remove them now since they lead to quite a bit of confusion, but at this point it would most likely cause too much code breakage in developers’ apps to remove them.

[Author]

Posts