Re: pitch tracking

[James Beauchamp <beaucham@xxxxxxxxxxxxxxxx>]

Charles Tassoni writes:

>Can anybody tell me how good the latest generation of pitchtrackers
>is?  Is there, for example, a pitchtracker that can take a melodic dictation
>from a human singer?  That is, if I sing a melodic sequence and two musicians
>produce the same transcription of that sequence, is there an algorithm that
>can als output that same transcription?
>        If accurate transcribers haven't been developed, what are some of the
>problems that have prevented their development?


Rob Maher and I have done some work in the pitchtracking area which we report 
in a paper coming out in JASA very soon (March?).  Judith Brown has published
several JASA papers on the subject.  In terms of converting the fundamental
frequency data to actual notes on staves, this could be done by converting
this data into a MIDI stream and then inputing it to a computer notation
program such as SCORE or FINALE. Some complete self-contained transcribers of
solo passages have been configured I'm sure. J.A. Moorer had a thesis on this
in the 70's (Stanford ~75) and Piszczalski and Galler at U. of Mich. developed
a computer system in the late 70's Lippold Haken had a real time system going
at the Univ. of Illinois CERL Music Group a few years back. There are probably
some commercial systems, as I hear that accompanying systems are available for
musicians.

I think none of these systems have been totally successful.  On rapid passages
there are lots of problems with interference from reverberation and/or various
noises of articulation during transitions between notes. The narrower the pitch
range, the easier the problem is to solve, but in music, you can have leaps of
two octaves or more. Octave errors are among the most common that occur in
these systems. With duets there are problems of separation, particularly when
harmonics of the two voices happen to collide.

There was a special session on music FF tracking at the fall 1992 ASA meeting
in New Orleans. No one had a definitive result, but there were lots of great
partial results. Maximum likelihood looks like a promising technique. Neural
nets may help.

Jim Beauchamp
UIUC