INTEGRATING LLM-BASED TOOLS INTO COMPUTER-ASSISTED COMPOSITION WORKFLOWS IN OPENMUSIC

1. OVERVIEW

This proposal presents a computer-assisted composition workflow that integrates Large Language Model (LLM)-based tools with the OpenMusic environment through the use of the Lisp programming language. In this approach, composers can generate functional code and design algorithmic musical structures by interacting with LLMs via natural language prompts, enhancing both accessibility and flexibility in contemporary compositional practices. However, it is essential to stress two points:

i) This workflow does not replace the creative act: the composer remains responsible for conceiving the core musical ideas that serve as the basis for algorithmic formalization;

ii) LLMs are still prone to frequent errors. Thus, a solid programming background remains necessary to critically evaluate, debug, and iteratively refine the code generated by the model.

In my presentation, I will illustrate this process by tracing the origins of my musical ideas — grounded in techniques and approaches I have studied from serialist and postserialist composers such as K. Stockhausen, F. Menezes, and B. Ferneyhough — and how I have adapted them to my own artistic language. I will present a recent composition that explores symbolic relationships between the poetics of John Cage, Clarice Lispector, and Zen Buddhism, drawing materials from texts, musical works, and sound recordings. For this piece, I developed a set of algorithms that enable transductions — or translations — between smooth time (real-time in seconds) and striated time (musical metric structures), as well as between text, musical notes, and durations. Additional algorithms allowed me to interrelate and translate heterogeneous data from texts, sounds, and musical structures, creating a compositional space where distinct formats and temporalities intersect.

Finally, I will reflect on the role of both my programming skills and limitations in shaping the creative process, and demonstrate how LLM tools were incorporated when appropriate, highlighting their potential and boundaries within artistic practice.

2. KEY COMPONENTS AND TECHNIQUES

I. Overall Structure
After introducing myself, the first part of this presentation will briefly introduce the composition that serves as the context for this research, outlining the conceptual and technical foundations that informed its design. In addition to the symbolic references to Cage, Clarice Lispector, and Zen philosophy, the piece draws heavily on historical techniques of material generation and formal structuring. These include early serialist procedures, such as Stockhausen's methods for control and material selection based on numeric profiles, as exemplified in Studie II and Gesang der Jünglinge; complex rhythmic manipulation strategies inspired by Bryan Ferneyhough, as discussed by Mikhail Malt in The Composition of Complex Rhythms (IRCAM, 2000); and pitch transformation techniques developed by Brazilian composer Flo Menezes, notably his concept of retroactive interlock. Together, these influences form the methodological basis from which the integration of LLMbased tools into the compositional process emerges.

II. Algorithmic Tools Development

The second part of this presentation will focus on a selection of algorithms and symbolic systems I designed specifically for this piece. These tools embody the
conceptual and structural principles discussed earlier, translating them into programmable, generative processes that operate within a computer-assisted composition workflow. 

The objects-functions I will present are:

a. Seq-Metric-Alea (Metric Sequence Generator)
Given a target duration in seconds and a base BPM, this algorithm
generates a random sequence of time signatures by combining
numerators and denominators from user-defined lists. The BPM
defines the tempo context for calculating the duration of each
measure. Through an adjustment process, the result is a valid
combination of measures that adds up to exactly the specified
total time, offering both structural control and aleatoric variability.

ii. Integer-Partition-Permutation
A method for generating rhythmic structures based on all possible
partitions of integers, combined with permutation strategies. This
approach allows for systematic yet varied generation of rhythmic
trees in OpenMusic, supporting a post-serial approach to rhythm
construction.

iii. Alpha->Note & Alpha->Interval (Symbolic Modes)
This system establishes correspondences between letters of the
alphabet and precise pitches/intervals using midicent values. It
offers six distinct modes for interpreting the alphabet, including
chromatic, microtonal, and irregular mappings derived from
subdivisions of the alphabet. These modes allow the integration of
linguistic material into musical structures, enabling words or
phrases to generate pitch sequences.

iv. Profile->abs-index
This system applies numerical profiles to control the selection
and reordering of elements within a dataset. It operates through
multiple modes that determine how profiles traverse or reorganize
the material, enabling both local permutations and global
structural variation. The method facilitates the systematic
generation of recurrent patterns, controlled randomness, or
formal development based on predefined sequences.

3. CONCLUSION
I will close this presentation by reflecting on the implications of incorporating LLMbased tools into compositional practice. While there is a clear risk that over-reliance on these models could discourage the effort required to formalize structures and understand the underlying code, there is also an alternative perspective. These tools, when approached critically, can function as interactive learning environments. Rather than replacing the composer's technical development, they can foster it.

This talk is part of IRCAM Forum Workshops Hors-les-Murs 2025 Rīga-Liepāja (Latvia)