Abstract
This paper presents an evolutionary computing approach to bridge the gap between mechanical, computer-generated music and the nuanced performances of human musicians. By utilizing a sequential covering genetic algorithm, the system automatically induces interpretable rules from professional Jazz saxophone recordings to predict how a performer deviates from a literal musical score. The model focuses on three primary expressive dimensions: timing (onsets), duration, and energy. Experimental results demonstrate that the evolved ruleset captures significant expressive intent, achieving high correlation coefficients with actual human performance data.
Introduction
When a computer plays a musical score literally, the result is often perceived as "robotic" or mechanical. Human performers inject expressivity by subtly altering the timing of note onsets, stretching or shortening note durations, and varying loudness (energy) based on the musical context. Traditionally, capturing this "intuition" required researchers to manually define rules using statistical analysis or rigid mathematical formulas. However, the complexity of professional performance makes manual modeling difficult.
This research addresses the challenge of automated rule induction. Instead of relying on human-defined heuristics, the proposed system "listens" to real-world performances and learns the underlying principles of expression through an evolutionary framework.
Methodology
The researchers developed a multi-stage pipeline to transform raw audio recordings into a generative model capable of predicting expressive deviations.
1. Feature Extraction and Symbolic Representation
The process begins with spectral analysis (Discrete Fourier Transform) of saxophone recordings to extract low-level descriptors such as instantaneous energy and fundamental frequency. To ensure robust pitch detection in potentially noisy environments, a "Two-Way Mismatch" (TWM) procedure is employed.
Once the audio is segmented into individual notes, each note is described by its context. This includes its relationship to neighboring notes (pitch and duration intervals) and its position within the musical structure. The system specifically incorporates Narmour’s Implication/Realization model, a musicological theory that describes how melodic patterns create and satisfy listener expectations.
2. The Sequential Covering Genetic Algorithm
The core of the learning engine is a sequential covering genetic algorithm. This approach follows a specific iterative logic:
- Rule Discovery: The algorithm searches for a rule that explains a specific subset of the performance data (e.g., identifying why a performer consistently lengthens notes in a particular metrical context).
- Data Covering: Once a rule is identified, the examples it successfully explains are removed from the dataset, and the algorithm repeats the process to find the next rule.
- Evolutionary Optimization: The rules themselves are evolved. A population of candidate rules, represented as bit-strings, undergoes selection, crossover, and mutation to find the most accurate descriptors of expressive behavior.
3. Magnitude Prediction
While the genetic algorithm classifies the type of change (e.g., "lengthen note"), a linear regression model is utilized to predict the magnitude of that change. This hybrid approach allows the system to determine both the logic behind an expressive choice and the intensity of its application.
Experimental Setup and Datasets
The study utilized a specialized dataset of monophonic Jazz recordings consisting of four standards: Body and Soul, Once I Loved, Like Someone in Love, and Up Jumped Spring.
A professional saxophone player performed each piece at 11 different tempos, resulting in a dataset of 4,360 performed notes. Each note was meticulously annotated with attributes including metrical strength, Narmour group classification, and tempo relative to the nominal speed.
Results and Insights
The model was evaluated using 10-fold cross-validation, measuring the correlation between the model’s predictions and the actual deviations made by the human performer. The results indicate a high level of predictive accuracy:
- Onset Deviation: 0.75 correlation
- Duration Transformation: 0.84 correlation
- Energy Variation: 0.86 correlation
Beyond numerical accuracy, a major advantage of this evolutionary approach is interpretability. Unlike "black box" neural networks, the system produces human-readable rules. For example, one evolved rule might state: "At nominal tempo, if the next note is similar and the current note is in a strong metrical position, then lengthen the current note." This transparency allows musicologists to analyze the learned behaviors to better understand the mechanics of performance.
Conclusion and Future Work
The sequential covering approach allows the system to "peel the onion" of musical expression—capturing broad, general principles first before identifying more niche, specific expressive exceptions.
However, limitations remain. Because the model was trained on a single professional saxophonist, the rules reflect a highly personal style that may not generalize to other performers. Additionally, the current system focuses on note-level transformations (start and end points) and does not yet account for "intra-note" expressivity, such as vibrato or dynamic swells within a single sustained note. Future research aims to expand the dataset to include diverse performers and incorporate timbre features—the "texture" of the sound—into the predictive model.
Summary updated