Automatic Phrase Continuation from Guitar and Bass-guitar Melodies

A framework is proposed for generating musically similar and interesting variations of a given monophonic melody. In this work, the focus is on rock/pop guitar and bass-guitar melodies with the aim of extensions to other instruments and musical styles. The original audio melody in audio format is first segmented into its component notes using onset detection and pitch estimation. Clustering is performed on the pitches to obtain a symbolic representation of pitch sequences in it. The note onsets are aligned with the estimated meter of the melody for a time-homogeneous symbolization of the rhythm in terms of onsets/rests and the metrical locations of their occurrence. A joint representation based on the cross-product of these two individual representations is used to train the prediction framework - the variable length Markov chain. It is hypothesized that such a model will rearrange, while maintaining certain coherence, the segments of the original melody. The musical quality of the generated melodies was evaluated through a questionnaire by a group of experts and received an overall positive response.