Cooperation between man and machine is an interesting area in the development of creative systems. It would be ideal if the system could generate new creative outputs, interpret them and discuss them with their contributors. A recent paper on arXiv.org presents a study on the generation of interpretations of musical decisions made by AI music systems.
The researchers leveraged the ability of transformer-based technologies to generate explanations of musical decisions using GPT-3, a state-of-the-art natural language model. It is provided with example tracks with explanations provided by the author. The ability to self-explain was then tested.
Experiments show that, in this context, GPT-3’s ability to learn several times is still limited. It can interpret musical notation but lacks the intelligence necessary to really understand it. The researchers concluded that there is a need to refine the model of musical interpretation.
Open AI’s language model, GPT-3, has shown great potential for many NLP tasks, with applications in many different fields. In this work, we perform the first study of GPT-3’s ability to communicate musical decisions through written interpretation when prompted with the written representation of a piece of music. music. Enabling dialogue in human-AI music partnerships is an important step toward more creative and engaging human-AI interactions. Our results suggest that GPT-3 lacks the intelligence necessary to truly understand musical decisions. A major barrier to better performance is the lack of data that includes explanations of artists’ creative processes for musical works. We believe such a resource will aid understanding and cooperation with AI music systems.
Research articles: Krol, SJ, Llano, MT and McCormack, J., “Towards the Generation of Music Interpreters with GPT-3”, 2022. Link: https://arxiv.org/abs/2206.08264
