Generating Abstract Rhythm Streams

This thesis presents the development of an accompaniment system capable of generating abstract rhythmic streams in real time. Building on the prior work of Behzad
Haki, particularly the GrooveTransformer generative system, our work shifts away from a drum-centric approach by focusing on rhythm as an abstract structure rather
than as instrument-specific events. The system is based on a transformer-based generative model trained on a newly curated dataset of diverse rhythmic material,
extending beyond strictly percussive sources.
To provide performers with greater flexibility, we introduce a set of novel control features that shape the output while preserving its rhythmic integrity. These include
mechanisms for adjusting rhythmic similarity, accent similarity, and other parameters designed for use in live improvisation contexts. By limiting the number of controls, the system is optimized for real-time performance, encouraging deliberate musical choices while leaving room for expressive interaction.
Our evaluation demonstrates that the system can successfully generate coherent and musically useful rhythmic streams in real time. However, further work is needed to
refine dataset curation and to redesign control features to improve usability and expressiveness in live settings. Ultimately, this research contributes to the field of  generative music by offering an approach to rhythm generation that emphasizes abstraction, performer agency, and real-time creative interaction.