TL;DR
This paper evaluates MMM-C, a creative AI system for music composition integrated into Cubase, by measuring its usability, user experience, and acceptance among expert musicians. The study found positive usability and acceptance scores, with users reporting novelty and ease of use, but also limitations in controllability and predictability, suggesting that a single-parameter interface may not be sufficient for expert co-creation.
What Problem Does This Paper Address?
The increasing prevalence of AI in creative domains, particularly music composition, necessitates a deeper understanding of how these systems impact human users. While many AI music generation systems focus on producing novel content, their integration into professional workflows and their effect on user experience, usability, and acceptance are less explored. This paper addresses this gap by evaluating MMM-C, a system designed for human-AI co-composition, specifically within the context of a popular Digital Audio Workstation (DAW) and with expert musicians. The goal is to understand the implications of such systems for musical practice and to inform the design of future co-creative AI tools.
Methods
The study employed a mixed-methods approach, combining quantitative and qualitative data collection. The core of the methodology involved a 3-part study where expert composers (hobbyists and professionals) interacted with MMM-C to complete three distinct music composition tasks: arrangement, variation, and original composition.
- Usability: Measured using the System Usability Scale (SUS) and the Creativity Support Index (CSI). A custom set of questions also assessed controllability.
- User Experience: Explored through open-ended qualitative questions focusing on trust, perceived quality, authorship, and flexibility.
- Acceptance: Evaluated using the Technology Acceptance Model (TAM), which assesses perceived usefulness and perceived ease of use.
The MMM-C system itself is a "1-parameter" plugin interface for the Multi-Track Music Machine (MMM) model, integrated into Cubase. Users could select musical content, adjust a "temperature" parameter (controlling conservatism vs. experimentation), and generate new musical sequences.
Datasets
Not explicitly detailed as training datasets for MMM-C within this paper. The paper mentions that MMM is trained on the MetaMIDI dataset, a collection of half a million MIDI files. For the study, participants were provided with existing genre-specific MIDI files (pop, rock, electronic, classical) or could use their own.
Key Insights and Ideas
- The "1-Parameter" Dilemma: While a single parameter (temperature) simplifies interaction, it significantly limits controllability and predictability for expert users, leading to a desire for more granular control over musical elements.
- Novelty and Surprise as Creative Drivers: Users reported positive experiences of novelty and surprise from the AI's output, which can be a valuable source of inspiration and help overcome creative blocks.
- Authorship Remains with the Human: Despite AI's generative capabilities, expert composers still feel a strong sense of creative ownership, viewing the AI as a tool for inspiration and curation rather than a replacement for their creative input.
- Expert vs. Hobbyist Similarities: Interestingly, no significant differences were found between hobbyist and professional expert groups in terms of usability, acceptance, or user experience, suggesting the core challenges and benefits of the system are consistent across these expert levels.
- Integration is Key: The integration of MMM-C into a familiar DAW like Cubase is crucial for its adoption and practical use by professional musicians.
Key Results
- Usability: Overall SUS scores were acceptable across all tasks (ranging from 71.43 to 75.71 out of 100). User-friendliness was rated positively, though task-based friendliness showed a slight dip for the original composition task.
- Controllability: Users expressed a strong desire for more control, with 9.54/10 indicating a need for more flexibility. Common themes included difficulty steering the system and a lack of parameters.
- Creativity Support Index (CSI): Enjoyment was the highest-scoring factor (3.85/5), followed by Exploration and Engagement (3.54/5). Expressiveness was the lowest (2.85/5), likely due to the limited control.
- Acceptance (TAM): Perceived Ease of Use scores ranged from 3.12 to 3.26 out of 5, and Perceived Usefulness scores ranged from 3.42 to 3.68 out of 5. While positive, these scores suggest that more work is needed for widespread adoption.
- Qualitative Findings: Users frequently mentioned "repeated generation" to find suitable outputs, highlighted the "non-deterministic" nature of the AI, and described "co-creation" through curating and editing AI-generated content.
Limitations and Future Work
- Limitations: The primary limitation identified is the restrictive "1-parameter" interface, which hinders controllability and predictability for expert users. The study also focused on expert composers, so findings may not generalize to novices.
- Future Work: The authors propose re-running the study with an interface that exposes more of MMM's controls (e.g., note polyphony, duration, density) to understand more complex parametric interactions. They also suggest adapting the methodology for beginner composers and exploring integration with other DAWs. Further research into the non-deterministic nature of AI systems and its impact on co-creation is also recommended.
Summary updated