[From Rick Marken (930808.1330)]
Bruce Nevin (Fri 930723 15:34:08 EDT) --
Why confine yourself to the point of view of a hearer who has never spoken?
When it comes to speech recognition, I tend to confine myself to the hearer,
who may or may not have ever spoken.
Rick, you don't have to have training as a linguist or a phonetician
to learn the relationship between speech sounds and what you have to
do to produce them.
No argument there.
The training is about naming them, categorizing
them, representing them by funny marks on paper, and so on, but none
of that is necessary for modelling what speakers and hearers do
unconsciously.
For modelling what speakers do -- yes. I question the importance of
articulation when modelling what HEARERS (recognizers) do. It just
seems like articulation is a useless step in the recognition process.
This seems so obvious to me. I agree that you have to learn how to
articulate (produce output variations) if you want to speak or play
music. But I don't think you have to use your articulatory control
systems if all you are doing is recognizing the sound. In fact, just
yesterday I recognized all three voices of Bach's C-minor fugue, though
I can only articulate one voice at a time (singing or playing the piano).
I can even recognize sounds that I can't articulate at all -- like
the speech generated by that guy who talks real fast in commercials.
A child who talks to you has indeed "learned the regular
relationship between phonemes and articulations", or she would not
be able to pronounce words so that you could recognize them.
I never said otherwise. This is certainly true of a child who talks.
It is just not (I argue) true of a child (or person) who recognizes
speech. Articulatory perceptions are important for production, NOT
recognition. I went to the dentist yesterday and had a large part of
my articulatory system anesthetized; but I was still able to understand
what the dentist was saying. Articulation MAY be involved in speech
recognition but I need to be convinced. The fact that acoustical variations
reflect underlying articulatory invariants does not convince me that
speech recognition is based on detecting the articulatory cause of
the acoustical signal; it seems AS IF this is happening but (for reasons
given in previous posts -- the main reason being that you would have to know
so much about the acoustical signal before applying the acticulatory analysis
that such an analysis would be superfluous) I don't believe that it is.
Maybe the best thing to do now is look at some of the evidence that
articulation is or is not required for speech recognition. Maybe this
is what Lance Norskog was referring to in his post of 090626 where
he said:
When you get tired of communing with your articulators,
there is a very fascinating treatment of all this in
a PhD thesis at CCRMA (Stanford's computer music lab).
Does this system control its perception of the results of its vocal
tract outputs? Does this system recognize acoustical inputs? How about
a brief description of the thesis?
Dag Forssell (930723 1520)--
Rick, I think you are denying the laws of nature as PCT describes
them, when you try to get people without first testing what their
"problems" are.
I presume that anyone doing science has a permanent "problem" to solve;
the problem of understanding how phenomena work. I assume that life
scientists want to understand how the phenomenon of behavior works
and I offer PCT as a possible solution. If they don't want it, that's
fine. This isn't "Glengarry Glen Ross" (a motion picture that gives a
"nice" depiction of my view of what selling is all about). I am not
interested in "selling" PCT; I am interested in showing the scientific
community some fundemental facts about the nature of behavior, control
and living control systems. If there is anything to these ideas (and,
obviously, I think there is) then science will eventually come around --
maybe in two days, maybe in two decades, but it will eventually come
around. I'm in no rush -- I'm not working on commission. And as long as
I have two or three brilliant colleagues to work with (like Bill P. and
Tom B.) then the work will be fun, if not profitable.
Best
Rick