Maximum entropy models capture melodic styles

Area of Science:

• Computational Musicology
• Machine Learning
• Digital Signal Processing

Background:

• Automatic music generation traditionally relies on Markov models, which can suffer from overfitting and require high-order interactions to capture long-range musical structures.
• Capturing the statistical nuances of musical melodies while avoiding plagiarism is a significant challenge in computational musicology.

Purpose of the Study:

• To develop a Maximum Entropy model for music generation that emulates the style of a given musical corpus.
• To demonstrate that complex musical phrases can emerge from simpler, pairwise interactions, reducing model complexity and overfitting.
• To validate the model's ability to generate novel yet stylistically consistent musical sequences.

Main Methods:

• Implemented a Maximum Entropy model utilizing a k-nearest neighbor approach with only pairwise interactions.
• Compared the proposed model against fixed-order and variable-order Markov models for melody generation.
• Employed a data-compression technique to quantitatively assess the originality and stylistic adherence of generated melodies.

Main Results:

• The Maximum Entropy model successfully captured the statistics of melodies, generating new sequences in the style of the input corpus.
• Long-range musical phrases emerged organically from competing pairwise interactions, negating the need for explicit high-order dependencies.
• The proposed model demonstrated superior performance over traditional Markov models in terms of stylistic emulation and innovation.

Conclusions:

• Maximum Entropy models based on pairwise interactions offer an effective alternative to high-order Markov models for music generation.
• This approach allows for the creation of musically sensible variations of existing phrases, fostering innovation in artificial music.
• The model provides a robust framework for generating novel musical content that balances stylistic fidelity with originality.