3
$\begingroup$

A set of gammatone band-pass filters are often used to model the filtering performed by the human inner ear. For example, "BatSLAM: Simultaneous Localization and Mapping Using Biomimetic Sonar" Jan Steckel and Herbert Peremans, as well as "Biologically inspired methods in speech recognition and synthesis: closing the loop" by Trevor Bekolay both use them. However, like all filters, if you start the input signal with silence, the gammatone filter responds differently than if you started the input signal with some sort of noise. Similarly, mammals are very rarely submerged in absolute silence. The environment and internal organs generate what I would imagine to be some sort of low frequency noise. Maybe it's a low-pass filtered white-noise?

If you are trying to use gammatone filter in a model, what do you prepend the start of your signal with? In other words, how are gammatone filters typically "primed"?

$\endgroup$
1
$\begingroup$

Mostly it depends on the kind of experiment you're doing. If you're interested in speech (as I was), most of the important information is in the 300-3000 Hz range, where low-frequency noise will have little effect unless it's extremely loud. If the experiment you're doing deals with static or other sounds with broad power spectra, then low frequency noise could have a significant effect.

As with most questions of perception, what matters most isn't really the signal itself but how the signal is changing. It's a big change to go from from silence to loud static. It's a lesser change to go from quiet static to loud static. That difference is more significant than the difference between going from silence to speech and going from quiet static to human speech, because speech is similar to neither silence nor static.

$\endgroup$
1
$\begingroup$

Priming the filters (or setting the initial state) is essentially unimportant for studies of audition. In a typical study, the signal is presented at 70 dB and the background noise of a quality double-walled sound both might be 20 dB. This gives a 50 dB SNR. That means if the amplitude of the filter response to the signal is 1, the response to the noise would be 10^-5 (0.00001).

If you are using the filter bank to study absolute detection, where the signal is small relative to the background noise, then you need to worry about the ongoing background noise (but should probably also include some sort of internal noise in the model). Even in these cases, the effect of priming will be small for signals of reasonable duration since the filters die out in a few milliseconds. So unless you are looking at absolute detection of brief signals, there is probably no reason to worry about "priming" the filter.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.