Music Genre Classification: A Semi-supervised Approach

Soujanya Poria, Alexander F. Gelbukh, Amir Hussain, Sivaji Bandyopadhyay, Newton Howard

Introduction

In the era of digital music streaming, accurate metadata is crucial for effective search, retrieval, and recommendation. However, manually labeling millions of tracks is prohibitively expensive and slow. A significant challenge in automated classification is the inherent ambiguity of music genres; many tracks share characteristics across multiple categories, such as the overlap between Rock and Metal. Traditional "hard" classifiers, which force a single label immediately, often fail to capture these nuances. This paper presents a high-performance music genre classification system that addresses these challenges by combining unsupervised fuzzy clustering with supervised Support Vector Machines (SVM) in a semi-supervised framework.

Proposed Methodology

The researchers developed a four-stage "fuzzy-to-hard" pipeline designed to leverage the strengths of both unsupervised and supervised learning:

  1. Feature Extraction: The system extracts three distinct types of audio features: short-time (timbral characteristics like MFCCs), long-time (statistical aggregates of timbre over several seconds), and beat features (rhythmic patterns and BPM).
  2. Fuzzy Clustering: Using a modified Fuzzy C-means algorithm, the dataset is grouped into 10 clusters. Unlike standard clustering, this "fuzzy" approach allows a track to maintain membership in multiple clusters simultaneously, reflecting the natural ambiguity of musical genres.
  3. Mapping: A small subset of labeled data is used via majority voting to map these fuzzy clusters to real-world genre labels.
  4. Hard Categorization (The Confusion Set Strategy): This novel stage narrows the search space. Instead of requiring an SVM to distinguish between all 10 genres, the fuzzy clustering results identify the top $K$ most likely candidates (the "confusion set"). A specialized SVM then performs the final classification between these few candidates.
Accuracy of various feature combinations across classification methods.

A key insight from the methodology is the importance of feature selection. As shown in Figure 1, "long-time" features—which capture the texture of sound over time—are significantly more effective for genre classification than instantaneous "short-time" features. The hybrid Fuzzy + SVM approach consistently outperforms standalone methods across all feature combinations.

Experimental Results

The study utilized the GTZAN dataset, a benchmark containing 1,000 balanced tracks across 10 genres. The proposed method achieved a remarkable 97.10% accuracy, significantly outperforming traditional standalone classifiers.

Performance comparison against baseline classifiers and sensitivity to the confusion set size K.

As illustrated in Figure 2, the system's performance is highly sensitive to the size of the confusion set ($K$). The optimal results were achieved at $K=2$, where the SVM only needs to distinguish between the two most likely genres identified by the fuzzy clustering. When compared to standard algorithms like K-Nearest Neighbors (54.21%), Naïve Bayes (65.88%), and Multi-Layer Perceptrons (74.23%), the proposed procedure offers a substantial leap in classification performance.

Conclusion and Future Work

By reducing the classification search space through fuzzy clustering, the researchers have demonstrated that a "fuzzy-to-hard" strategy can achieve near-perfect accuracy on benchmark datasets. While the results are promising, future work is needed to scale the system for industrial applications involving millions of tracks, where the computational cost of training multiple binary SVMs may become a factor. Additionally, the authors suggest that incorporating non-audio data, such as lyrics and emotional metadata, could further refine classification in real-world, noisy environments.

Summary updated

Abstract

Music genres can be seen as categorical descriptions used to classify music basing on various characteristics such as instrumentation, pitch, rhythmic structure, and harmonic contents. Automatic music genre classification is important for music retrieval in large music collections on the web. We build a classifier that learns from very few labeled examples plus a large quantity of unlabeled data, and show that our methodology outperforms existing supervised and unsupervised approaches. We also identify salient features useful for music genre classification. We achieve 97.1% accuracy of 10-way classification on real-world audio collections.
BibTeX
@inproceedings{poria2013music,
  author = {Soujanya Poria and Alexander F. Gelbukh and Amir Hussain and Sivaji Bandyopadhyay and Newton Howard},
  title = {{Music Genre Classification: A Semi-supervised Approach}},
  booktitle = {MCPR},
  year = {2013},
  doi = {10.1007/978-3-642-38989-4\_26},
}
Figure