A Pattern Recognition Approach for Melody Track Selection in MIDI Files.

摘要: Abstract Standard MIDI files contain data that can be considered asa symbolic representation of music (a digital score), andmost them are structured as a number tracks. One ofthem usually contains the melodic line piece, whilethe other tracks accompaniment music. The goal ofthis work is to identify track melody us-ing statistical properties musical content and patternrecognition techniques. Finding very useful fora applications, like speeding up match-ing when searching in databases or motif extraction,among others. First, set descriptors from each ofthe target file extracted. These inputto random forest classifier assigns probability ofbeing track. with high-est selected one containing melodicline file. Promising results have been obtainedtesting different styles.Keywords: Melody finding, analysis, rep-resentation, perception.