[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: AUDITORY Digest - 4 Mar 2008 to 5 Mar 2008 (#2008-52)



Dear Stuart,

I am wondering why you suggest that the off and on diag elements of the CM might not be different.

Point 1:
They are different, since the 1-diag element is the average of the off-diag elements. In fact all the information of interest is in the off-diag elements, and each diag element is meaningless (but it is stable, since it is an average of lots of very different things). Each diag elements is due to very different types of confusions.


Furthermore, the average over diag elements (P_c), is really meaningless.!!!

Taking averages of very different classes (uncorrelated things) is rarely a good idea.

Point 2:

phonetic features are not a function of the SNR, however the confusions P_{heard|spoken}(SNR) are. Therefore phonetic features (e.g., voicing etc.) do not appear to be useful for explaining the CM(SNR) data. We have found averages over phonetic features to be a less-than-useful concept. In fact, I believe this is a classic mistake of the literature, ever since JFH's monograph was published. Miller and Nicely 1955 further propagated this error.

Do you disagree?

I trust you are well. I am sorry I did not get a chance to meet with you when I was in London at Kings College for a month this last summer. I will be back in July for the Mech of hearing meeting in Keele. Maybe then?

Jont


REFS:

,author={Allen, Jont B.}
,title={Consonant recognition and the articulation index}
,journal=JASA
,year=2005
,month=apr
,volume=117
,number=4
,pages={2212-2223}


@article{PhatakAllen07a ,author={Phatak, S. and Allen, Jont B.} ,title={Consonant and vowel confusions in speech-weighted noise} ,journal=JASA ,year=2007 ,month=apr ,volume={121} ,number={4} ,pages={2312-26}


AUDITORY automatic digest system wrote:


Date: Wed, 5 Mar 2008 07:14:40 +0000 From: Stuart Rosen <stuart@xxxxxxxxxxxxxxxxxxx> Subject: Re: stability in CMs

I think you could take the approach of using log-linear modeling to
statistically compare the two matrices to one another, rather than
treating diagonals and off-diagonals as something different.

With this approach, you could also collapse the matrices by phonetic
features (voicing, place and manner) before comparing, so as to
determine if changes were limited to a particular feature.

This might be relevant:
Bell, T.S., Dirks, D.D., Levitt, H., & Dubno, J.R. .  Log-linear
modeling of consonant confusion data.  Journal of the Acoustical Society
of America, 79(2), 518-525.

Yours - Stuart Rosen

Yoon Yang-soo wrote:
 > Dear List,
 >
 > I'm working on analyses of confusion matrices measured twice from the
 > same group of HI listeners 6 month interval.
 >
 > I want to assess the stability of performance on both diagonals and
 > off-diagonls as a function of SNR and of test-retest.
 >
 > So far I found two relevant literatures (Dubno Dirks, 1982, Bilger
 > and Wang, 1976), reporting correlations between test and retest.
 >
 > I wonder if any knows ways to show the stability other than
 > correlations, and
 >
 > I wonder if anyone knows references other than those.