Subject: From: Pierre Divenyi <marva4!EarLab.UUCP!pierre(at)UCDAVIS.EDU> Date: Fri, 18 Feb 1994 11:18:25 UTC
marva4!ucdavis!VM1.MCGILL.CA!AUDITORY Subject: Re: How much precision needed? Message-Id: <761410196(at)EarLab> Date: Wed, 16 Feb 94 14:49:56 UTC From: marva4!EarLab.UUCP!pierre(at)ucdavis.edu (Pierre Divenyi) Reducing the number of channels (provided that the width of the channels will be inversely related to their number, as it should if we are to cover the same frequency range) will sacrifice frequency resolution in favor of time resolution. Since tracking envelope fluctuations is one of the primary processes needed for speech recognition, Bob Shannon's results make perfect sense. Plomp would be happy to hear about them. As far as Ed Burns's tongue-in cheek remark is concerned, obviously, the ear has the large number of channels required for processing a whole orchestra. However, the time-frequency trade-off still plays (no pun): for transients (percussion, attacks) the "what" will always matter less than the "when", even in music. Similarly, to resolve vowels in a vowel-rich language (such as Swedish, for example), one needs longer durations, i.e., to sacrifice time resolution. It is the number of bits that appears to be fixed, only the distribution of the bits varies. Gabor got the Nobel Prize for working out the details of this trade-off. (Personally, I like better Brillouin's treatise of the problem, for reasons of readability.) Pierre Divenyi PS: My ARO highlights were listening to Jim Simmons's bat IC firings, and the dinner at Brunello's (on the expensive side but, on the other hand, I had all my breakfasts in my room, courtesy of Publix).