Subject: Re: comodulation release of masking (CMR) From: at <parIPO.TUE.NL> Date: Mon, 5 Oct 1998 10:03:07 +0200Dear Neil, > >I think there is no consensus about the processing of CMR > >stimuli although we recently argued that the envelope cross correlation > >(not to mistake with the envelope cross covariance or correlation > >coefficient) may be helpful in accounting for CMR data > >(van de Par and Kohlrausch, 1998a/b: JASA 103 pp 3605-3620; 1573-1579). > >In the paper on page 3605 we even describe a model that accounts > >for CMR data. > > > >Steven > > > > > > I think we are agreed that some form of cross-correlation mechanism is > probably a good horse to bet on, I wonder though if you could clarify > your distinction between different forms of cross-correlation. The difference between both forms of correlation can be seen looking at their definitions: Cross correlation: rho = <xy> / sqrt(<x^2> <y^2>) where x and y are the envelopes for which one wants to obtain the envelope cross-correlation. Cross covariance: r = <XY> / sqrt(<X^2> <Y^2>) where X = x - <x> and Y = y - <y> (Note that the cross covariance is also called the correlation coefficient) In other words: for the cross covariance the mean (or DC) component of the envelopes is first removed (the -<x> and -<y> terms in the equation). This removal of the DC component is what is problematic about the cross covariance in my opinion. To see this, consider two tonal carriers at different frequencies that are modulated with the same sinusoidal modulator except that the modulators differ in phase by either zero or pi radians. Now subjects have to detect the change in modulator phase. Independent of the modulation depth the envelope cross covariance will either be +1 or -1, indicating that detectability of modulator phase difference is independent of modulation depth. This is a problem because eventually, when the modulation depth is very small subjects will have great difficulties detecting modulator phase differences. The envelope cross correlation doesn't suffer from these problems because the envelope cross correlation will depend on the modulation depth. The difference between the two definitions of correlation has also been studied in relation to binaural detection at high frequencies where processing of envelopes is also assumed to be important. A very interesting study has been published by Bernstein and Trahiotis on this topic (1996; JASA 100, 1754-1763) where behavioural data show that the envelope cross correlation gives a much better account of high frequency binaural detection data then the covariance. > The > cross-correlation mechanism I proposed in Todd (1996, Network: Computation > in Neural Systems. 7, 349-356) was a product-moment on the cosine phase > of the envelope modulation power spectrum. If I read this correctly you calculate correlations between envelope spectra instead of correlations between temporal envelopes. In that case the DC component of the envelope will be preserved within the spectrum. As far as I can see there will be no problems with regard to the type of stimuli that I described above. Best wishes, Steven van de Par IPO-Center for research on user-system interaction Den Dolech 2 5612 AZ Eindhoven The Netherlands Phone: +31 40 2475215 Fax: +31 40 2431930 E-mail: par(at)ipo.tue.nl > This had the advantage that > one didn't need to have delay lines or some other storage mechanism since > if one uses acausal impulse response function (or non-linear phase response > transfer function) for the modulation filter, this is effectively a kind of > memory. Further, the cosine phase spectrum locks into that of the envelope > thus preserving sensitivity to phase effects in streaming (e.g. in an > alternating A B A B sequence). I did play around with some other metrics, > e.g. Euclidean distance, but did not conclude that there was any advantage > over the product-moment, although I did consider how such a cross-correlation > mechanism might be instantiated neurally. > > Best wishes > > Neil > > University of Manchester > Manchester > M13 9PL > UK > Tel. +44 (0)161 275 2557 > McGill is running a new version of LISTSERV (1.8d on Windows NT). Information is available on the WEB at http://www.mcgill.ca/cc/listserv