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

Re: Intelligibility of reversed speech, Why?

Tóth László wrote:

> On Thu, 25 Jan 2001, Ward Drennan wrote:

> > Acoustic information in speech is redundant, of course, but certainly,
> > as information (a cue of any sort) was removed or otherwise
> > blurred, you would expect the intelligibility to decrease for a given
> > S/N ratio.
> >
> No. Redundancy means that under a given condition (e.g. S/N ratio) you
> don't necessarily need ALL the cues for perfect comprehension. That is,
> you can remove some cues without decreasing intelligibility. However,
> under some other conditions the same cue may be vital.
> To put it another way: how can you tell that something you removed from
> the signal was or wasn't a cue? My point was that you cannot state this
> for sure from examining its effect under only one possible condition, put
> you have to examine all possible conditions.
> I hope this helps enlight my thinking.

I have another problem with the question of redundancy in speech.

It is clear there is redundancy in once sense. Language is an entropy constrained
code. If I say "the dog bit the c_t's tail" I doubt anybody would have trouble
telling me that the missing letter is "a." Thus speech is redundant.

However I dont find the above to be very helpful, or enlightened. One should
not be testing with high-context low-entropy sentences. One should be
using nonsense words, or other ways to remove such context effects.
Miller, Heise and Licten (1950's) did this by showing the subjects a list that contained
the word they were about to hear, and this allowed them to use real words,
yet remove the entropy problem.

When one uses high entropy words, then I really wonder if speech is redundant.
My feeling is that it is not. This feeling is based on the 50 or so years of articulation
testing by fletcher and the gang. See my IEEE  review paper on this topic. (I also have
one in JASA.)

One problem with redundancy is that what do you do when the two redundant cues
give you different messages? In other words, they are telling you there is a problem.
This allows one to detect that the SNR is poor, but it just creates conflict. It doesn't
reduce the error rate.


Jont B. Allen
AT&T Labs-Research, Shannon Laboratory, E161
180 Park Ave., Florham Park NJ, 07932-0971
973/360-8545voice, x7111fax, http://www.research.att.com/~jba