A Framework for Computational Stylistic Analysis in Jazz Improvisation

Thesis Abstract

This study addresses the growing need for objective, replicable analysis of stylistic variation in jazz improvisation through a computational framework that integrates symbolic, acoustic, and contextual features. The aim is to develop and validate a principled framework for measuring stylistic distance and identifying signature motifs across performers and ensembles. Specific objectives include (1) to operationalize a multi-level feature space encompassing melodic, harmonic, rhythmic, timbral, and interactional cues; (2) to implement an interpretable machine learning pipeline that maps feature patterns to stylistic typologies; (3) to evaluate the framework’s discriminative power across soloists, ensembles, and historical periods; (4) to test the theoretical integration of cognitive-perceptual and social-signaling perspectives in improvisational stylistics; and (5) to propose a practical toolbox for researchers and educators. Methodologically, the study adopts a mixed-methods research design grounded in computational musicology and jazz theory. The population comprises commercially and academically released jazz performances from 1950 to 2020, with a purposive sample of 60 sextets and quartets representing diverse sub-genres (bebop, hard bop, modal, fusion) and notable improvisers (e.g., Parker, Coltrane, Miles Davis, Herbie Hancock). A stratified random sample of 120 improvisational solos and 60 ensemble passes will be annotated using a dual-track instrument (a) a symbolic feature extractor that captures note-level and chord-scale relationships, motifs, and transitional probabilities; and (b) an audio feature processor that derives timbre, dynamics, and groove-related descriptors from high-fidelity stems. Data collection employs existing transcriptions and high-quality audio, supplemented by expert-guided corrections for melodic integrity and rhythmic alignment. The data analysis employs a layered approach. First, descriptive statistics summarize feature distributions and autocorrelations to establish baseline stylistic landscapes. Second, a sequence modeling component uses Hidden Markov Models and Long Short-Term Memory networks to capture temporal dependencies in melodic trajectories and motif development. Third, a supervised learning framework—comprising random forests and gradient boosting—produces stylistic typologies, with evaluation via cross-validated accuracy, precision, recall, and F1 scores. Fourth, regression analyses (multivariate linear and logistic) examine the influence of performer identity, ensemble configuration, and historical period on stylistic metrics. Finally, a sociocultural interpretation draws on schemas from Bourdieu’s habitus and cognitive ergonomics to relate perceived style to social signaling and mental categorization processes. The framework’s validity is tested through triangulation with expert panel ratings (n=20 jazz scholars) and a subset of perceptual experiments (n=40 participants) assessing perceived similarity and intelligibility of improvised passages. Expected findings include (a) robust, interpretable feature sets that differentiate soloing strategies across artists and periods, (b) evidence that rhythmic and motif-trajectory features contribute most to stylistic discrimination, (c) measurable alignment between computationally derived typologies and expert-perceived stylistic categories, and (d) demonstrable utility of the framework in predicting audience-perceived authenticity and authorship. The study anticipates revealing nuanced interactions between individual improvisational choices and ensemble-level dynamics, with implications for pedagogy and music information retrieval. The study contributes to knowledge by delivering an integrated, transparent framework that unites symbolic and acoustic analyses with cognitive and social theories to quantify jazz improvisation style. It provides a validated pipeline and an open-source toolkit enabling replication, extension to other genres, and incorporation into music education curricula and performance analytics. The main conclusion is that computable stylistic signatures in jazz improvisation can be reliably extracted and interpreted through a multi-level model that respects musical structure while acknowledging social-contextual factors. Recommendations include expanding the sample to include non-Western jazz traditions, integrating real-time analysis for live performance settings, and developing user-friendly visualization modules to support educators and researchers.