An Order-Complexity Model for Aesthetic Quality Assessment of Symbolic Homophony Music Scores

Xin Jin, Wu Zhou, Jinyu Wang, Duo Xu, Yiqing Rong, Shuai Cui

Introduction

As artificial intelligence continues to advance in the creative arts, the evaluation of AI-generated music remains a significant challenge. While current models can generate technically proficient MIDI files, the results often feel "soulless"—characterized by monotony and a lack of the structural nuance found in human compositions. Traditionally, assessing the quality of these works has relied on subjective human listening tests, which are time-consuming and difficult to scale.

This research addresses the gap by proposing an objective, quantitative Score Aesthetic Assessment Model (SAAM). Specifically designed for symbolic homophony (music featuring a clear melody and accompaniment), the model utilizes music theory and information science to distinguish high-quality human compositions from lower-quality AI outputs with high precision.

Theoretical Framework: Birkhoff’s Aesthetic Measure

The foundation of the SAAM is grounded in Birkhoff’s Aesthetic Measure, a classical mathematical formula that defines beauty ($M$) as the ratio of Order ($O$) to Complexity ($C$):

$$M = \frac{O}{C}$$

To adapt this formula for music theory, the researchers decomposed Order and Complexity into four measurable high-level features:

  1. Order ($O$):
    • Harmony: Quantified through interval ratios and chord progression tension.
  2. Symmetry: Measured via self-similarity and the "skewness" of pitch and rhythm distributions.
  3. Complexity ($C$):
    • Shannon Entropy: Measures the uncertainty or chaos within the pitch and rhythmic sequences.
  4. Kolmogorov Complexity: Evaluates how much the music can be compressed; highly redundant music is considered "simpler" or less complex.

Methodology

The researchers developed a multi-stage pipeline to process and score music files. Using the music21 toolkit, the system extracts raw pitch, rhythm, and chord attributes. These are processed into eight basic features, which are then fed into four logistic regression layers to calculate the high-level aesthetic features (Harmony, Symmetry, Entropy, and K-Complexity).

Finally, these features are combined into the Order/Complexity ratio. A Sigmoid function maps the resulting value to a score between 0 and 1. This allows the model to classify a score as either "Good" (human-like aesthetic quality) or "Bad" (monotonous or low-quality AI generation).

Experimental Setup and Datasets

To validate the model, the study utilized a dataset designed to highlight the contrast between aesthetic and non-aesthetic music:

  • Positive Samples (Human): 100 pop song scores from Musescore. The researchers focused on score-based data rather than performance-based data (like the POP909 dataset) to ensure the evaluation was based on the composition itself rather than the skill of a performer.
  • Negative Samples (AI): 100 scores generated by Google Magenta’s improv_rnn. To ensure a fair comparison, the AI was constrained to the same chord progressions and key signatures as the human samples.

The dataset was split 7:3 for training and testing purposes.

Key Insights and Results

The model demonstrated exceptional performance in distinguishing human compositions from AI-generated ones, achieving 93.3% precision and a 90.9% F1-measure.

The Importance of Harmony

Ablation studies—where individual features were removed to test their impact—revealed that Harmony is the most critical component of the model. Without the Harmony feature, the model's ability to distinguish human from AI music (AUC) dropped from 0.93 to 0.77. This suggests that the structural relationships between notes and chords are the primary indicators of aesthetic quality in homophonic music.

The "Entropy Paradox"

One of the most interesting findings was that human scores often exhibited higher entropy (more uncertainty) than AI scores. While AI models tend to stay "safe" by hovering around the tonic (the home note), human composers introduce intentional variations and deviations that add a "pleasant" complexity. Furthermore, AI-generated music frequently lacked symmetry; human composers use repetition and self-similarity to create structural "rhyme," whereas AI sequences often appeared too random or lacked long-term structural cohesion.

Human Validation

To verify the model's assumptions, the researchers conducted a Turing test with 15 volunteers. The participants correctly identified human-composed music 91.3% of the time and rated it as "more beautiful" in 87% of cases, confirming that the model’s quantitative metrics align closely with human subjective judgment.

Conclusion and Future Work

The SAAM provides a robust framework for the objective evaluation of music, proving that a holistic approach combining music theory and information science is necessary to define aesthetic value. However, the authors note that the model currently measures adherence to "beautiful" rules rather than creativity—the intentional breaking of rules that often characterizes masterpieces.

Future research aims to expand the model's scope beyond homophony to include polyphonic music and complex genres like jazz or classical. Additionally, the researchers suggest that applying the Order-Complexity logic to raw audio files, rather than just symbolic scores, represents a promising next step for the field of automated music assessment.

Summary updated

BibTeX
@inproceedings{jin2023order,
  author = {Xin Jin and Wu Zhou and Jinyu Wang and Duo Xu and Yiqing Rong and Shuai Cui},
  title = {{An Order-Complexity Model for Aesthetic Quality Assessment of Symbolic Homophony Music Scores}},
  booktitle = {ICME},
  year = {2023},
  doi = {10.1109/icme55011.2023.00224},
}
Figure