Deep Improv
Overview
Deep Improv is an ongoing ensemble-based improvisation project centred on deep listening, collective performance, and continuous recording. Every rehearsal and performance is captured in full, forming a growing archive of musical interaction that is used as primary material for analysis and creative development.
Rather than focusing on collecting MIDI or symbolic representations, the project works directly with audio. This includes live instrumental performance, electronically generated sound, and a library of samples, many drawn from public domain orchestral recordings. These materials are treated as fluid sources for transformation, recombination, and reinterpretation during playing.
The recorded archive is analysed using deep learning frameworks centred on representation learning and latent structure inference. In place of explicit, hand-crafted feature extraction, the approach relies on autoencoding architectures, nonlinear dimensionality reduction, and sequence models to learn from the audio stream.
Particular emphasis is placed on aligned-dimensional embedding (ADE) strategies, in which temporally distributed events are mapped into a shared latent space and examined through graph-based formalisations. We also use Catalan-type combinatorial structures that are proving to provide a useful lens for modelling hierarchical segmentation and dependencies across musical time. The aim is to understand the more subtle relationships between gesture, texture, and ensemble interaction.
These extracted structures are then fed into a real-time synthesis environment built in SuperCollider, which acts as another member of the ensemble, responding to learned patterns, tendencies, and emergent behaviours within the group. The result is what we loosely termed “smart improv”, where human performers and machine processes co-exist within the same musical space.
Ultimately, the idea is to explore what happens when listening is extended beyond the moment of performance — when the traces of improvisation are recursively analysed, transformed, and reintroduced into the creative process.
Background Reading
Background reading and related literature can be found in the background reading archive. This includes work in machine learning, computational creativity, and experimental music practices that inform the development of the project.
Related Code
The codebase includes tools for large-scale audio recording, dataset organisation, feature extraction, and integration with SuperCollider. There is also ongoing work on deep learning pipelines for analysing rehearsal archives and generating responsive sonic material.
Getting Involved
We sometimes do group jams on a Jamulus server and will post details of these sessions here. If you are interested in experimental improvisation, the project, or participating in future sessions, you are welcome to get in touch.