Search for a command to run...
• Generative model for piano sight-reading exercises controlled by difficulty level. • Scores are generated in MusicXML, ensuring readability and playability. • An auxiliary loss improves difficulty control and prevents conditioning collapse. • Expert pianists rated the pieces as clear, natural, and suitable for teaching. • The approach advances towards personalized music learning. Sight-reading is a core skill in music education. It refers to the ability to play a written piece of music correctly the first time it is seen. Developing this skill requires frequent practice with completely new musical excerpts that match the difficulty level of the student. However, creating new sight-reading exercises at a specific difficulty level requires significant time and expert knowledge. As a result, students and teachers often rely on pieces from the existing piano literature, even though sight-reading exams typically use compositions written specifically for the exam. Generative music systems provide a promising approach for creating new sight-reading material with explicit control over performance difficulty. In this work, we frame the creation of sight-reading exercises as a symbolic music generation task that produces piano scores with controllable difficulty. Existing approaches typically rely on control tokens to guide generation, but we show that this strategy does not result in piano scores with reliably controlled difficulty. To address this issue, we introduce an auxiliary difficulty prediction objective using synthetic difficulty labels produced by an expert-based system, enabling scalable training. Our method improves difficulty conditioning accuracy from 69.3% to 92.9% compared to a baseline that conditions generation solely on difficulty control tokens, and reduces mean squared error from 0.30 to 0.09. A user study with expert pianists shows that the generated scores are rated comparable or superior to human-written material in readability and naturalness, while maintaining appropriate playability across difficulty levels. These results represent a step toward the generation of music exercises for a variety of educational applications.