Abstract
Cover song detection, a core task in Music Information Retrieval (MIR), often relies on Query-By-Example (QBE) systems that face a performance "glass ceiling." Traditional post-processing methods like clustering are often brittle, requiring manually tuned similarity thresholds. This paper introduces the Spectral Connectivity Network (SCN), a parameter-free, unsupervised approach that applies spectral graph theory to refine similarity scores. By modeling the music collection as a graph and utilizing a Green’s function approach to weight eigenvectors, SCN achieves significant accuracy gains, particularly for lower-fidelity systems, without the need for explicit cluster count definitions.
Introduction
In QBE systems, the goal is to retrieve all versions (covers) of a query song from a large database. While initial similarity engines provide a baseline, they often produce noisy results where the distributions of cover and non-cover pairs overlap significantly.
To improve these results, researchers have turned to post-processing techniques that exploit the transitive relationships between songs (e.g., if Song A is similar to Song B and Song B to Song C, then A and C are likely covers). However, standard methods like UPGMA clustering are highly sensitive to parameters like the similarity threshold. If the dataset or the underlying algorithm changes, these parameters often require tedious re-tuning. SCN addresses this by providing a global, parameter-free framework.
The Spectral Connectivity Network (SCN)
The proposed SCN method treats the entire music collection as a graph where songs are nodes and similarity scores are edges. The process moves beyond local pair-wise comparisons to a global analysis of the network structure.
Methodology
- Similarity Transformation: Raw dissimilarity scores are converted into a similarity matrix using a radial basis function (RBF) with local scaling.
- Laplacian Normalization: The matrix is normalized to account for node degrees, mitigating the influence of "hub" songs that might otherwise cause false positives.
- Eigenvalue Decomposition: The system performs an eigenvalue decomposition on the normalized Laplacian.
- Green’s Function Integration: Unlike standard spectral clustering which requires choosing a fixed number of eigenvectors ($k$), SCN employs a Green’s function. This approach weights eigenvectors by their corresponding eigenvalues, allowing the system to capture the global data structure without knowing the number of cover sets in advance.
Experimental Evaluation
The authors tested SCN across three environments: synthetic datasets with controlled noise, a large real-world collection of 2,125 songs, and the benchmark Covers80 dataset.
Key Results
- Robustness to Noise: SCN's most significant advantage is its performance in high-noise environments. In simulations, while traditional community clustering performance degraded with noise, SCN remained stable.
- Performance Gains: On the Covers80 dataset using a lower-performing base algorithm, SCN achieved a 19.70% improvement in Mean Average Precision (MAP), more than doubling the gain seen with UPGMA (8.06%).
- High-Fidelity Systems: For high-performing initial engines, SCN provided a consistent 1-3% relative MAP increase, proving its utility even when the baseline is strong.
Conclusion
The Spectral Connectivity Network offers a robust, parameter-free alternative to traditional clustering for cover song detection. By shifting from local thresholding to global connectivity analysis, it effectively breaks the performance glass ceiling of QBE systems. Future work will focus on optimizing eigenvector selection and improving scalability for million-song databases.
Summary updated