Score-based Style Recognition Using Artificial Neural Networks

Overview The original idea was to develop a system for musicological analysis that was capable of assisting in the resolution of issues concerning compositional authenticity. Based on explicit rule-based interrogation of a musical score the system gathers statistical information by way of a data extraction engine, the subsequent neural network implicitly forms an abstract impression of habitual characteristics within the composition. It must not be assumed that this system aims towards the modeling of human musical perception, as it is the authorÕs belief that score-based analyses cannot adequately meet this task. Initially the system was developed to explore the authenticity of flute compositions attributed to Frederick II ÒThe GreatÓ. Since there has long been musicological debate concerning this matter it was decided to acquire information from both performers of period instruments and musicologists concerning the characteristics and signatures required in the discrimination task. Based on free and multiple-choice reports returned a rule base was compiled that was then implemented into the system. Methodology For the purpose of this study it was decided to create a corpus of score representations in ALMA 1 , an antiquated and relatively forgotten format. The primary reason for this was the ease of coding without the necessity to sacrifice any of the printed score attributes. Since the system is based on intervallic difference of note relations it was not necessary to transpose the selected material to a common key-signature, which appears to be common practice in similar studies. Aside from the standard frequency of note distribution statistics, horizontal pitch-class ÒsnapshotsÓ are used in order to obtain a rough image of the tonal contents of each measure. Other data is collected from functions stemming from Lerdahl & Jackendoff (1983) theories on tonal music. In a similar manner vertical pitch-class analysis is employed, this method involves the weighting of individual events based on their metric position to obtain a single vector. Weighting of importance of the vertical pitch-class data is crucial for the perceived tonality of each measure (Cook 2000). Strong-weak interplay between parts provides major individualistic cues in the identification process. Auxiliary and passing notes in this instance are not considered since the resolution is restricted to minimal values of 16 th notes. Efforts to locate modulations were developed by monitoring the frequency of note occurrence over time. Having established the initial key a scan is run over the score to track any deviation, which is …