An SVM-based classification approach to musical audio

Namunu Chinthaka Maddage, Changsheng Xu, Ye WangISMIR 2003 · OpenReview

Abstract

This paper presents a hierarchical classification framework for musical audio using Support Vector Machines (SVM). By categorizing audio into vocal, instrumental, and mixed tracks, the system addresses the complexities of overlapping signal structures. A key contribution is the Correction Algorithm for Music Sequence (CAMS), which utilizes rhythmic constraints to refine temporal classification. Experimental results demonstrate that the SVM-based approach, enhanced by CAMS, achieves up to 95.78% accuracy, outperforming traditional Gaussian Mixture Models and Neural Networks.

Introduction

The rapid growth of digital music libraries has necessitated robust automated content-based retrieval systems. Musical audio classification is challenging due to the temporal and spectral overlap of vocal and instrumental components. This study evaluates the efficacy of SVMs in a hierarchical structure to distinguish between vocal melodies, specific instruments, and their mixtures.

Methods

The proposed system employs a multi-stage hierarchical tree of SVM classifiers to process audio data systematically.

Hierarchical SVM classification tree for musical audio.

Feature extraction involves analyzing 20ms frames for Linear Prediction Cepstral Coefficients (LPCC), Mel-Frequency Cepstral Coefficients (MFCC), Spectral Power (SP), and Zero Crossing Rates (ZC). The hierarchy begins by separating pure vocal or instrumental tracks from mixed audio, followed by finer classifications such as gender detection and instrument identification (e.g., acoustic guitar vs. harmonica).

A critical component of the methodology is the Correction Algorithm for Music Sequence (CAMS). Standard frame-by-frame classification often suffers from temporal instability or "flickering" errors. CAMS leverages the inherent rhythmic logic of music—specifically the beat frequency—to ensure that classified segments meet a minimum duration threshold. Short, likely erroneous segments are reassigned to neighboring classes based on Euclidean distance, effectively denoising the classification sequence.

Results

The SVM-based approach consistently outperformed Multilayer Neural Networks (MNN) and Gaussian Mixture Models (GMM) across all tested categories.

Comparison of classifier performance with and without the CAMS algorithm.

Baseline SVM accuracy for broad classifications reached approximately 92.8%. The integration of the CAMS algorithm provided a significant performance boost of 2% to 4%, leading to a peak accuracy of 95.78% for distinguishing vocal music from mixed music. Sub-classification of instruments and gender within mixed tracks also maintained high accuracy levels, often exceeding 94%.

Conclusion

The study highlights the power of combining hierarchical SVMs with rhythm-aware post-processing. By treating music as a structured time-series rather than isolated frames, the CAMS algorithm significantly improves classification reliability. While the current results on a controlled dataset of Sri Lankan music are promising, future work will focus on scaling the system to larger, more diverse datasets and optimizing computational efficiency through dimensionality reduction.

Summary updated

BibTeX
@inproceedings{maddage2003svm,
  author = {Namunu Chinthaka Maddage and Changsheng Xu and Ye Wang},
  title = {{An SVM-based classification approach to musical audio}},
  booktitle = {ISMIR},
  year = {2003},
  url = {https://openreview.net/forum?id=oWqi-YX3U3J},
}
Figure