Abstract:

This paper presents an in-depth analysis of MusicGen, a generative music transformer model, focusing on the capabilities of its self-attention heads in understanding and representing diverse musical elements. We uncover how MusicGen encodes various aspects of music through head-wise probing, from instrument recognition to more complex downstream tasks. Our findings reveal that certain attention heads are particularly adept at discerning specific musical characteristics, suggesting a pathway to highly nuanced music generation. By leveraging techniques for inference-time control, originally developed for large language models, we discuss the potential for achieving additional precise control in text-to-music generation tasks. This approach allows for fine-grained customization beyond basic text prompts, facilitating music generation that more accurately reflects the user’s creative intent.