Adaptive, real-time generative accompaniment system for solo performers, 2024–2026

Thesis Abstract

This study addresses the challenge of creating responsive, musically coherent accompaniment for solo performers in real-time, leveraging computational creativity to bridge performer intention and automated sonic texture. The problem centers on achieving tightly synchronized timing, idiomatic harmonic progressions, and expressive dynamism in live contexts where latency, interpretive variability, and genre familiarity constrain conventional accompaniment systems. The aim is to design, implement, and evaluate an adaptive, real-time generative accompaniment system that learns performer style and context, and responds with temporally aligned, musically appropriate material across diverse repertoires. Specific objectives are (1) to develop a modular system architecture combining a predictive timing engine, a conditional generative model for harmony and texture, and an expressive control interface; (2) to implement low-latency inference pipelines enabling sub-100 ms cueing for accompaniment decisions; (3) to integrate Bayesian learning and reinforcement learning components to adapt to performer preferences over a 12-week usage period; (4) to evaluate system efficacy across three genres (classical, contemporary, and film-inspired scores) using both expert musicians and trained performers as participants; (5) to assess perceptual quality, timing accuracy, and perceived agency via mixed-methods analysis. The methodology adopts a design, implementation, and evaluation framework grounded in human-computer interaction and music information retrieval theories. The research design combines iterative prototyping with a quasi-longitudinal field study. The population comprises professional pianists and violinists (N = 18) and student performers (N = 12), recruited from two conservatories, with a target sample ensuring representation across stylistic preferences. Data collection instruments include (a) instrumented performance recordings under three repertoire conditions (predefined classical piece, improvised contemporary piece, and genre-blind excerpts), (b) system interaction logs capturing latency, decision confidence, and control input, (c) perceptual evaluation scales including the Music Similarity Rating and the Geneva Implicit Agency Scale, and (d) semi-structured interviews post-session. The data collection spans a 6-month period with 10 session blocks per participant, totaling approximately 4200 performance minutes and 360 hours of systemlog data. Validity and reliability are addressed through pilot testing (n=6) of the control interface, test-retest reliability checks for perceptual scales, and triangulation across quantitative and qualitative data. Analytical approaches combine quantitative and qualitative methods. Latency and timing accuracy will be analyzed using time-alignment error metrics and mixed-effects ANOVA to assess effects of repertoire type and performer experience. Harmony and texture appropriateness will be examined through computational musicology measures (tonal centroid stability, chordal surprisal, rhythmic irregularity) and assessed via multivariate regression to determine predictors of perceived musical coherence. The system’s adaptation dynamics will be modeled with hierarchical Bayesian inference to quantify learning rates and personalization effects across performers. The latent structure of improvisational support will be explored via topic modeling of session transcripts and thematic analysis of interview data, triangulated with system usage metrics to identify factors driving perceived agency and musical alignment. A conceptually grounded evaluation will compare the adaptive system against a baseline non-adaptive accompaniment in a within-subject design, employing paired t-tests and effect size calculations. Expected findings indicate that the adaptive system delivers significantly lower timing deviations (mean improvement of 12–18 ms) and higher coherence scores in harmony and texture continuity, particularly in contemporary and improvised contexts. The Bayesian adaptation component is anticipated to demonstrate rapid convergence of performer-specific models within 4–6 sessions, with measurable improvements in perceived agency and satisfaction. The interpretive analyses are expected to reveal nuanced preferences for harmonic color and atmospheric texture, moderated by performer experience and repertoire. The study contributes to knowledge by operationalizing a scalable architecture for real-time generative accompaniment that integrates predictive timing, controllable expressive parameters, and performer-centric adaptation within live performance conditions. It extends theoretical discussions in computational creativity, human–computer collaboration in music, and embodied interaction. Practical implications include a validated framework for designing stage-ready accompaniment systems, guidelines for latency budgeting, and actionable insights into interface design for performer control. The main conclusion posits that tightly integrated real-time generative models, when attuned to performer style through Bayesian learning and reinforced by perceptual validation, can substantially enhance the sense of musical collaboration and expressivity in solo performance. Recommendations address extensions to multi-instrument ensembles, cross-cultural repertoires, and deployment in performance venues with networked audio ecosystems, as well as avenues for refining evaluative instruments to capture nuanced perceptual and cognitive load aspects.