Unsupervised Generation of Chord Sequences from a Sound Example

In this thesis project a system is developed for the analysis of a chord sequence given as an audio input with the aim of generating arbitrarily long musically meaningful and interesting sound sequence using the input characteristics. The procedure that is followed includes the transcription of a harmonic sequence into a shued multilevel representation, utilizing a tempo estimation procedure and identifying the most regular subsequence in order to guarantee that the specic harmonic structure is preserved in the generated sequence. In the final synthesis, the system recombines the audio material derived from the sample itself and it is able to learn various concepts of harmony and individual styles from the data. Also it can show how practice deviates from the ideal form. First of all, the sound is split into chord segments, then a clustering model is applied for grouping the chords, taking into account their harmonic structure through their Constant Q Proles. Following this procedure, a Variable Lenght Marcov Chain (VLMC) model is used in order to predict and re-shuffle these elements, maximizing simultaneously the cluster resolution, in order to avoid musical discontinuities. The system is evaluated objectively as well as subjectively by musicians and non-musicians, showing that the automatically generated chord sequences maintain the key features of the original and that they can be musically interesting.