Leonardo Lesmo, Vincenzo Lombardo, Rossana Damiano, Daniele Radicioni
Fingering in Music Performance
Our research addresses the modeling of the music performance process,
that is the transformation of symbolic representations of a score into
physical gestures, needed to operate a music instrument: a model of music
consists of the interpretation of the score, and the application of the gestures to some sound synthesis device that represents the instrument.
Musical performance is a challenging field of research, involving highly
specialized cognitive and motoral skills; performance modeling benefits
from several knowledge sources, as the knowledge upon the piece and the
score, and about the technical aspects of the instrument for which the
score has been conceived, the composer's intentions and the execution
style. All these issues contribute to fingering, which therefore involves
several competences, such as
musical analysis, for the interpretation of the notes in input, physical constraints, posed by the instrument where the notes have to be played, bio-mechanical constraints, which characterize the possible figures of the hand.
Based on motor-behavioral and psychological evidences, we propose a computational
model of fingering for string instruments, which encapsulates the main
physical and bio-mechanical constraints coming from the performer's hand
and from the instrument. The model performs a path-search in a graph representing all the possible fingerings where dynamic-programming features are coupled with a CSP formalization of some gestural patterns.
Automatic Music Analysis
This line of research aims at exploring how to perform automatic music analysis: we address both music "vertical" dimension (that is, harmony), and music "horizontal" dimension (that is, melody). Namely, underpinned by evidences from literature, we argue that music analysis can be represented as a particular case of sequences analysis.
Hence we cast the problem to a Supervised Sequential Learning (SSL) task.
SSL problem can be solved through many techniques, such as the state-of-art
algorithm HMPerceptron. After having collected 2 corpora of analyzed musical
pieces, we have devised a set of boolean features adequate for music analysis.
These features provide enough discriminative power to the HMPerceptron implemented to solve the sequential learning problem: experimentation provided encouraging results.