CAUI Demonstration — Composing Music Based on Human Feelings

AAAI 2002

TL;DR

This paper presents a novel system that automatically composes music tailored to evoke specific human feelings. By learning the relationship between musical structures and subjective emotional responses, the system can generate personalized musical pieces, moving beyond traditional music generation methods.

What Problem Does This Paper Address?

The creation of music that reliably evokes specific emotions in listeners is a complex challenge. While music theory provides general guidelines, individual emotional responses to music are highly subjective and influenced by personal experiences and cultural context. Existing music composition systems often rely on predefined rules or stylistic imitation, lacking the ability to adapt to individual emotional preferences. This paper addresses the gap by proposing a system that learns and models these subjective emotional responses to musical structures, enabling the automatic generation of music designed to elicit a desired feeling in a specific individual.

Methods

The core of the system lies in its ability to learn the relationship between musical structures and human feelings. This is achieved through a two-pronged approach:

  • Musical Structure Extraction: The system analyzes musical pieces to extract key structural elements, specifically focusing on melody and chord progressions. Chords are categorized by their harmonic function (e.g., tonic, dominant, subdominant). The system learns rules for assigning chords to melodies based on sequences of these functions.
  • Personalized Feeling Model: The system constructs a "personal model" of a user's feelings by collecting their emotional evaluations of various musical pieces. Users rate pieces using adjective pairs (e.g., bright-dark, happy-unhappy). Machine learning techniques, specifically inductive learning, are used to derive constraints that map musical structures to these subjective feelings.

Once a personal model is established, the system can:

  • Automatic Arrangement: Rearrange existing musical pieces to better match the user's desired emotional response.
  • Automatic Composition: Generate entirely new musical pieces from scratch using a melody generator guided by a chord progression derived from a genetic algorithm, which uses the personal feeling model as its fitness function.

The system is described as a "Constructive Adaptive User Interface" (CAUI), emphasizing its ability to build new descriptions adaptively based on user interaction and learned models.

Datasets

The paper mentions the use of:

  • 75 well-known music pieces without modulation: These were used for extracting musical structures and training the system.
  • 3 original pieces: These were used for automatic arrangement experiments.
  • Japanese JPOP songs and pieces from classic music or textbooks for harmonics: A total of 39 + 36 = 75 pieces were used for harmonic analysis.

The paper does not explicitly detail a single, named dataset for evaluation but rather describes the experimental setup where subjects evaluated arranged pieces.

Key Insights and Ideas

  • Subjectivity of Musical Emotion: The paper highlights that emotional responses to music are not universal but highly personal, necessitating individualized models.
  • Musical Structure as a Predictor of Emotion: By identifying specific musical structures (melody and chord functions), the system demonstrates that these elements can be reliably linked to perceived emotions.
  • Machine Learning for Personalization: Inductive learning and genetic algorithms are powerful tools for building personalized models of musical preference and generating novel content that aligns with these models.
  • Adaptive User Interfaces Beyond Simple Adaptation: The CAUI concept extends adaptive interfaces by not just adapting to user behavior but actively constructing new representations (musical pieces) based on learned user models.

Key Results

The paper presents qualitative results in the form of two created pieces (Figure 3 and Figure 4) that were designed to be favorite pieces for specific subjects (A and B). The authors claim these examples demonstrate the system's ability to compose favorite pieces without prior knowledge of the subject's favorite songs, by automatically acquiring their preferred musical structures. They also mention that the system produced 36 arranged pieces for each subject in their experiments, suggesting a significant output of personalized music.

Limitations and Future Work

  • Limited Scope of Musical Elements: The current system primarily focuses on melody and chords. The authors acknowledge that other musical elements (e.g., rhythm, timbre, dynamics) also contribute significantly to emotional perception and could be incorporated.
  • Generalizability of Personal Models: While the system aims to create personalized models, the extent to which these models can be generalized across different musical genres or styles is not fully explored.
  • Subjectivity in Evaluation: The reliance on subjective user evaluations for training and validation introduces inherent variability.
  • Future Work: The authors suggest extending the system to incorporate more musical elements and exploring more sophisticated machine learning techniques for improved prediction and generation. They also imply that the CAUI framework can be applied to other domains beyond music composition.

Summary updated

BibTeX
@inproceedings{2002caui,
  title = {{CAUI Demonstration — Composing Music Based on Human Feelings}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2002},
}
Figure