Abstract
We introduce CONLON, a pattern-based MIDI generation method that employs a new lossless pianoroll-like data description in which velocities and durations are stored in separate channels. CONLON uses Wasserstein autoencoders as the underlying generative model. Its generation strategy is similar to interpolation, where MIDI pseudo-songs are obtained by concatenating patterns decoded from smooth trajectories in the embedding space, but aims to produce a smooth result in the pattern space by computing optimal trajectories as the solution of a widest-path problem. A set of surveys enrolling 69 professional musicians shows that our system, when trained on datasets of carefully selected and coherent patterns, is able to produce pseudo-songs that are musically consistent and potentially useful for professional musicians. Additional materials can be found at https://paolo-f.github.io/CONLON/ .
Original language | English |
---|---|
Title of host publication | Proceedings of the 21st International Society for Music Information Retrieval Conference |
Editors | Julie Cummings, Jin Ha Lee, Brian McFee, Markus Schedl, Johanna Devaney, Corey McKay, Eva Zangerle, Timothy de Reuse |
Publisher | International Society for Music Information Retrieval |
Pages | 876-883 |
Publication status | Published - Oct 2020 |
Event | 21st International Society for Music Information Retrieval (virtual) - Duration: 11 Oct 2020 → 16 Oct 2020 |
Conference
Conference | 21st International Society for Music Information Retrieval (virtual) |
---|---|
Period | 11/10/20 → 16/10/20 |
Keywords
- Machine learning
- Music generation
- Wasserstein autoencoders
- Pianoroll representation
- Interpolation