Learning Jazz Pianist Style with Cross-Attention Conditioning (almostimplemented.github.io)

A new AI project has successfully trained a model named Aria to emulate the playing styles of fifteen legendary jazz pianists, inspired by Dick Hyman's 1994 work that captures their unique signatures through original compositions. By fine-tuning Aria, a 16-layer transformer pretrained on piano MIDI, the researchers introduced a gated cross-attention layer, allowing the model to condition its output based on the learned embeddings of specific pianists. This innovative approach resulted in an impressive 70% attribution accuracy when identifying the pianist's style in generated music segments, significantly outperforming the 37% accuracy achieved without conditioning. This development is significant for the AI/ML community as it demonstrates a novel method for generating music that authentically reflects the unique characteristics of individual artists. By highlighting how the model's conditioning affects musical accuracy and style transfer—measured through a classifier designed to attribute generated pieces to their intended pianists—the research opens up new avenues for AI in creative fields, such as music composition and performance. The study emphasizes the importance of controlled experimentation in generative AI, setting the stage for further exploration into the intersections of AI, art, and human creativity.