Reproduce “BachProp: Learning to Compose Music in Multiple Styles” on OpenResearch

About this paper

Florian Colombo, Wulfram Gerstner

Researchers at EPFL developed BachProp, an algorithmic composer that employs a novel (time-shift, duration, pitch) representation and a deep GRU network to generate music across diverse styles. This approach achieved high prediction accuracies, such as 97% for time-shifts on Bach Chorales, and yielded compositions that human evaluators rated as highly as, or even preferred over, original human-composed pieces.

Key insights

• The normalized (time-shift, duration, pitch) representation allows a single model architecture to effectively learn and generate music across a wide range of styles without requiring specific structural assumptions.
• Sequential, conditional prediction of note features (timing, duration, pitch) enables the network to capture intricate musical dependencies, resulting in more coherent and musically structured compositions.
• Subjective evaluations demonstrated that AI-generated music, when composed with this method, can achieve perceptual quality comparable to, and sometimes preferred over, human-composed originals, pushing the boundaries of computational creativity.

Summary by alphaXiv.