Path: utzoo!utgpu!jarvis.csri.toronto.edu!mailrus!accuvax.nwu.edu!ferret!sandell
From: sandell@ferret.ils.nwu.edu (Greg Sandell)
Newsgroups: comp.music
Subject: Re: programs that can infer key/meter
Keywords: Finale?
Message-ID: <1277@accuvax.nwu.edu>
Date: 13 Oct 89 18:40:56 GMT
References: <15170@netnews.upenn.edu> <125936@sun.Eng.Sun.COM> <7203@thor.acc.stolaf.edu> <484@artsnet.UUCP> <7354@thor.acc.stolaf.edu>
Sender: news@accuvax.nwu.edu
Reply-To: sandell@ferret (Greg Sandell)
Organization: ^
Lines: 98

For some reason, my response got posted as a null posting.  Here's another
try....

Mark Gresham writes:
>But competing musical elements are as often the norm as they are
>the exception.

Competing musical elements are often local ambiguities that are 
resolved at higher levels.  One gestalt construct that L & J uses
is the "theory of good continuation" which says that if the 
piece has had previous indications of being in 3/4 time, then it
probably will stay in 3/4 time.  If an ambiguity is localized to
a small number of measures, and the preceding and following
portions suggest 3/4, and 3/4 can continue through the passage
correctly on a correct-number-of-beats level, then the theory is
biased to assume that 3/4 stayed in effect.

> The
>'strong' accents (thinking only metrically, even) do not
>necessarily fall on the first pulse of any given measure.

L&J distinguish between 'phenomenal' accents (accents overtly cued
by the presence of a note and reinforced by, say, dynamic markings)
and 'metrical' accents.  Metric accents are expectations that are
set up by other non-ambiguous cues.  Although there may be silence
on the first beat, metrical expectations cause a listener to 
experience it as a strong beat anyway.

>>Of course, real music is more complicated than this simple example, but the book
>>also describes more complicated conditions.
>
>But there's more problems with simple conditions:
>  (various metric ambiguities cited...)

I used the wrong phrase; I meant that there are additional, more complex
rules to handle the richness of "real" music.  Many classic ambiguities
such as hemiola rhythm are handled by rules in the system.  The system
would be able to handle other ambiguities it had never encountered
before, provided there were other cues for the meter, which there
frequently are.  But it would be easy to try to systematically
confound the approach by removing all cues.  When you get to that
point, even the listener has trouble coming up with a "correct meter."
And in any case, what L&J are trying to do is to model the
perceptions of an experienced listener.

>Even an example from the sixteenth century, "Ecce quomodo moritur"
>(Jacob Handl) would be made into a metrical hash by the
>Lerdahl/Jackendoff approach to analysis

Why would you want to 'discover the meter' of music of this period?
I don't know the specific piece, but I recognize the genre.  Even
if written in 4/4 time nobody *hears* a 4/4 meter marching along
in this music.  In any case, L&J limit their scope to music of
the common practice period.  (Complain if you like, but you try to
create an exhaustive theory of all musics of all periods, and see
how much success you have!)
>
>> (Stuff on the Krumhansl key-finding approach here...)

>Greg, most music of an oral tradition contradicts that, with the
>dominant being the most *frequently occurring* note, followed by
>the mediant, then the submediant; the same goes for spontaneous songs
>made by children.

I'm aware of this; in Irish folk music, for example, tunes frequently
end on scale-degree 5.  The algorithm might work in folk genres anyway,
because even if there are pieces which have more mediant and submediant
than tonic (really?), the tonic notes which are indeed present will
probably fall at metrically important locations.  (Recall that I
said the algorithm's note count is *weighted* by surface
emphases of notes, such as metric position, dynamics.) If you don't even
have this condition, then I submit that you have a key-ambiguous piece,
and in fact the algorithm would find more than one key vying for "first
place."  (The algorithm outputs a strength match with all 24 major and
minor keys, and the strongest key, assuming that one really stands out,
is selected as the answer.)

Mark, your questions are interesting, and your challanges are merited.  But
I think maybe you thought that what L&J are trying to do is discover
the *notated* meter.  To the contrary, they proceed from the assumption
that the notated meter is frequently in contrast to the experienced
meter.  On the Krumhansl algorithm, its success rate speaks for itself.
Except for the weighting mechanisms, it is an unmusical approach, so I
was expecting to get a rise out of somebody on the net when I brought
it up!

>
>Cheers,
>
>--Mark

Greg


***************************************************************
* Greg Sandell, Institute for Learning Sciences, Evanston, IL *
* sandell@ferret.ils.nwu.edu                                  *
***************************************************************