Generating Melodies from Melodic Outlines Towards an Improvisation Support Systems for Non-musicians

Generating Melodies from Melodic Outlines Towards an Improvisation Support Systems for Non-musicians Tetsuro Kitahara Nihon University, Japan Twitter: @tetsurokitahara

Our goal To enable everyone to enjoy music improvisation including non-musicians Cmaj7 Am7 Create a melody Play it Chord progression Harmony theory Music knowledge Musical scale Do-Re-Mi Learned musical phrases

Our previous work: JamSketch User draws: System generates: Melodic outline Demo Video Intuitive, easy to draw Melody based on a genetic algorithm

Today's talk Re-developed the melody generation method using a CNN Chord time Output Melody (notes) Melodic outline Input time

Issue in designing a CNN Music is not shift-invariant Pitch axis Time axis Rarely appears in the C-major key Different notes have different roles The most important note on-beat off-beat On-beats and off-beats have different features

Filter design in our CNN model filter size (avoid shift in the pitch axis) conv. conv. filter size pitch Melodic outline ch ch Chord progression time avoid that on-beats and off-beats are flipped deconv. deconv. ch pitch time time Pitch (cont) Pitch time time

Dataset Input data Output data Melodic outlines made by smoothing melodies (a sequence of notes) 96 Blues melodies from Weimar Jazz Database Half for training, the rest for testing

Demo https://bit.ly/3zB2Tja

• https://bit.ly/3zB2Tja

Evaluation: Method Melodic outline Melody from DB (given melody) Model Generated melody MP3 MP3 https://bit.ly/3tmDt5a Folder 30 persons having musical experience Rate 8 melodies randomly chosen (1–7) based on 6 criteria

• https://bit.ly/3tmDt5a

Evaluation: Results Obtained close scores to given melodies

Conclusion JamSketch: an improvisation system where the user draws melodic outlines Developed the melody generation method with a CNN Input Melodic outline Listening test Output Melody Implies close quality to human melodies