Rhythmic Explorations in Latent Spaces (RELS) is a browser-based musical performance led by an interrogation on the creative affordances that machine learning can provide in musical performance practice. Specifically, the performer is interested in the expressive possibilities that a machine learning-driven musical instrument can contribute to a performer when playing with rhythmic patterns. In the performance for the Network Music Festival 2020, a custom-built web-based interface is used to explore latent spaces learned from rhythmic patterns of footwork, 2-step, and gqom.
These rhythms are particularly interesting to play with since they can be perceived at their tempo, or at their tempo octave (i.e., twice as fast or slow). This characteristic enables performers to play with two different tempi at once, leaving the audience free to embody the music as they see fit with their mood and context. These rhythms are also interesting because they use binary and ternary metric grid overlaid one of top of the other, leading to more musical variability.
A dataset of drum beats of the mentioned genres is used as training data. Then a used RhythmVAE, a variational autoencoder neural network topology, is used to learn a compressed representation of the patterns. This is ported to the browser as a playback version of the network so that the performer can map the browser canvas to the high-dimensional rhythmic space. Then, the latent space is expored by dragging an imaginary playback head that allows retrieva;, interpolation, and extrapolation of points in the space, generating variations of the original rhythms.
Vigliensoni, aka Gabriel Vigliensoni, is a Montréal-based musician and producer from Chile. His artistic work is informed by formal musical training and extensive studies in sound recording, music production, new musical interfaces, and music information retrieval. He is currently a postdoctoral research fellow at Goldsmiths, University of London, doing practice-based research on the creative capabilities and affordances of the deep learning paradigm and applying it to assisting musical composition.
In his work, Vigliensoni explores the different stages of the music production workflow, always seeking to transform the process of making a record into a playground for learning and experimentation. Throughout his career he has experimented with techno and breakbeat, overlapped krautrock and electronica, explored vocal-driven songs that eschew the standard pop format, used procedural composition techniques, and relied on extended structures, slowing down beats and doing live drum programming to bring liveness and immediacy to the workflow of digital music production.