1. Introduction

The Pitch Track compares melody features on a SOM.

2. Requirements

2.1 Only melody pieces

In order to extract a meaningful melody feature, the pieces under consideration must contain music that exihibits meldoy. Drum grooves and the like will be sorted out. Furthermore, we take only unision pieces. Hence, we check for harmonicity by computing the correlation dimension with constant frequency. If their mean is below a threshold the piece is passed on for further analyses. To bypass this process, ESRA should provide an check button, for unison pieces, which can be checked by users during upload.

2.2 f_0 estimation

f_0 is estimated as strongest autocorrelation peak with lag > 1.

2.3 Scale agnostic

The next analysis step requires all melodies to live in the same scale. Hence, we need to estimate the scale. We treat the most common pitch class per song as the root.

2.4 n-gram

Trigrams are computed from each pitch sequence. The ten most commemon trigrams and their respective frquency are taken to form a melody feature vector.

2.5 Comparison

We compare the feature vectors element-wise, that is, we correlate the most commen trigrams, the second most common, and so on. The correlation function to be used depends on musical view point and is not yet specified.