Detecting background noise level

[zhousean]

Hi Halle,

I am using a recognitionScore threshold to skip some inaccurate hypotheses. I found the recognitionScore varies depends on the background noise level. In a noisy environment the recognitionScore will be a larger number whereas in a quiet environment the recognitionScore will be a smaller number.

I was wondering whether there is any delegate methods in OpenEars that allow me to detect the noise level and then I will dynamically set the recognitionScore threshold in my app.

Thanks for your help as always.

[Halle Winkler]

Hello,

Sorry, you can’t use scores for anything except relative comparison within the same session while it is in similar environmental circumstances. They can’t be used to evaluate across sessions or across speakers since they are sensitive to many factors including distance, noise, accent, nearness of the speaking frequency to the mean that was used in the model, the device, and the vocabulary size and complexity. You will end up with misleading data if you try to use scores to isolate only one of those factors.

[zhousean]

Thanks Halle.

It we cannot use score, is there anyway to dynamically recalibrate OpenEars to the changing background?

[Halle Winkler]

It is dynamically recalibrating itself, so you can’t recalibrate it more than it is already doing. Have you tried using the VoiceChat audio mode?

[zhousean]

No, I haven’t tried the VoiceChat audio mode. I actually don’t know anything about this mode. Could you give a bit more details about it?

[Halle Winkler]

Hi,

There is some documentation in the OEPocketsphinxController header – do you have specific questions that aren’t addressed there?

[Author]

Posts