Abstract
Identifying structural boundaries in music—such as transitions between verses and choruses—is a foundational task in music information retrieval. This paper presents a hybrid framework that integrates supervised and unsupervised learning techniques to improve boundary detection. By combining mid-level features like dissonance and vocal occurrence with a graph-theoretic approach that incorporates temporal priors for section lengths, the system achieves a significant performance boost over traditional single-method approaches.
Introduction
Music boundary detection enables essential features like automated summarization and structure-aware playback. However, the task is notoriously difficult. Unsupervised methods, which rely on repetition or clustering, often struggle to differentiate between distinct sections that consistently appear together. Conversely, supervised methods trained on acoustic features can be rigid and fail to generalize to songs outside the training distribution. This research seeks to bridge this gap by developing a system that is both data-driven and song-adaptive.
Methodology
The proposed system operates at the beat level, utilizing the BeatRoot algorithm to identify beat onsets, as musical boundaries typically align with these points. The pipeline consists of feature extraction, dual-path estimation (supervised and unsupervised), and a final integration stage.
Feature Engineering
Beyond standard low-level features like MFCCs and chromagrams, the authors introduce three critical mid-level features to capture semantic and structural cues:
- Novelty Score: Quantifies local dissimilarity between adjacent segments using a checkerboard kernel on a similarity matrix. This was found to be the most effective individual predictor.
- Dissonance Level: Based on music theory, this feature tracks "tension and release" by calculating the weighted sum of dissonant intervals. Boundaries often follow peaks in dissonance.
- Vocal Occurrence: An SVM-based classifier detects singing. Transitions often coincide with the entry or exit of a vocalist, providing a strong semantic cue.
Integrated Framework
The core of the system is a graph-theoretic fusion of different estimation models. Each beat onset is represented as a node in a directed graph.
- Vertex Weights: These represent the probability of a beat being a boundary, calculated by fusing outputs from a supervised SVM and an unsupervised clustering model. This "best of both worlds" approach provides both general discriminative power and song-specific context.
- Edge Weights: These incorporate "music knowledge" by modeling the probability of section lengths (e.g., 8 or 16 bars) based on training data statistics. This helps prune false positives that acoustic models might pick up.
- Optimization: The Viterbi algorithm is employed to find the optimal path through the graph, identifying the most likely sequence of boundaries for the entire song.
Experimental Results
The framework was evaluated on the RWC Music Dataset, comprising 100 Japanese and American pop songs. Using stratified five-fold cross-validation, the integrated approach demonstrated significant improvements over standalone methods:
- Supervised Only: F-score of 0.2641
- Unsupervised Only: F-score of 0.3517
- Integrated Approach: F-score of 0.4094
The results indicate that while unsupervised methods provide strong song-specific context, the fusion with supervised models and temporal priors significantly enhances precision and recall.
Conclusion
This research demonstrates that a hybrid approach—combining acoustic novelty, musicological features like dissonance, and temporal constraints—outperforms single-paradigm methods for music boundary detection. While the system still faces challenges with over-segmentation, future work will focus on incorporating more complex music knowledge into the graph and refining mid-level features to better distinguish between minor phrase pauses and major section transitions.
Summary updated