Musical style recognition.

[Pino Buzzanca <g_buzzanca@xxxxxxxxxxx>]

Dear List,
I'm using a back-propagation neural network to recognize musical style. The
idea is to provide the network with several vectors, each vector
representing a musical phrase. Training test consists of several musical
phrases belonging to 5 different composers (who all lived in the 17th
century).
Input layer consists of neurons with Vector Input (out of four input units
for each note, two encode rhythmic information, and two encode pitch
information).
I'm now concerned with the necessity to account for musical input of varying
length in an even manner: in other words the problem is that the network's
structure requires me to encode musical phrases in fixed-width
representations (!) and this is not realistic in comparison to real life (I
mean the 'amputation' of musical stimuli into fixed-width sequences, cfr.
Sloboda 1985).
Any ideas on segmentation possibilities that seem plausible from the
cognitive/auditory/methodological standpoint?
Thanks a lot in advance, Giuseppe Buzzanca.






==========================================
Giuseppe Buzzanca
Professor of Music
State Conservatory of Music
Via Brigata Bari, 26
70124 Bari ITALY
tel. +39 (080) 574 0022
fax +39 (080) 579 4461
g_buzzanca@virgilio.it