If Turing Played Piano With an Artificial Partner

Area of Science:

• Artificial Intelligence
• Music Technology
• Human-Computer Interaction

Background:

• Music is a social activity, yet AI development for interactive musical experiences lags behind score generation.
• Current AI models excel at producing music but struggle with real-time social interaction.
• Effective musical collaboration requires complementary contributions and adaptation between partners.

Purpose of the Study:

• To assess if a passive generative music model can enable social interaction with a human piano player.
• To quantify the user experience and performance quality of human-AI musical collaboration.
• To identify AI configurations that best support interactive musical experiences.

Main Methods:

• A variational autoencoder pretrained on piano scores was adapted for a timed call-and-response task.
• Human participants played piano with human partners and AI partners with varied parameters.
• Post-trial ratings assessed performance quality and self-other integration.

Main Results:

• AI partners were adequate for basic practice but not convincing for improvisation.
• AI configurations with simpler architectures and higher similarity to the human partner received the highest ratings.
• Simple, similar AI partners were not statistically different from human partners in realism, ease of interaction, and integration.

Conclusions:

• Basic AI musical partners can provide adequate practice experiences.
• Simplicity and similarity are key factors for successful human-AI musical interaction.
• Future AI partners may require principles from Artificial Life for genuinely interactive experiences.