Introduction
Music genre classification remains a challenging task in music information retrieval due to the subjective and often overlapping nature of genre boundaries. Traditional approaches often struggle to capture the complex spectral and temporal modulations that define musical styles. This paper presents a robust framework that leverages psycho-physiologically grounded audio features—specifically 3D cortical representations—and introduces a novel dimensionality reduction technique, Locality Preserving Non-Negative Tensor Factorization (LPNTF), to enhance classification performance using Sparse Representation-based Classification (SRC).
Methodology
The proposed framework follows a three-stage pipeline: feature extraction, dimensionality reduction, and classification.
Feature Extraction
Audio clips are mapped into 3D cortical representations using a computational auditory model. These tensors capture slow spectral and temporal modulations across scales, rates, and frequency bands, mimicking the human auditory cortex's processing.
Locality Preserving Non-Negative Tensor Factorization (LPNTF)
The primary innovation is the LPNTF algorithm. Standard tensor factorization techniques like Non-Negative Tensor Factorization (NTF) often ignore the intrinsic geometric structure of the data. LPNTF addresses this by incorporating a graph-based regularizer. By constructing a nearest-neighbor graph, LPNTF ensures that audio tracks close in the original high-dimensional space remain close in the reduced tensor space, preserving the data's manifold structure.
Sparse Representation-based Classification (SRC)
The reduced tensors are classified using SRC, where a test sample is modeled as a sparse linear combination of training atoms. The genre is determined by the class that minimizes the reconstruction error. The authors demonstrate that aggressive dimensionality reduction is essential for making SRC computationally feasible for high-dimensional audio data.
Experimental Setup
The framework was evaluated on two benchmark datasets:
- GTZAN: 1,000 recordings across 10 genres.
- ISMIR 2004 Genre: 1,458 recordings across 6 genres.
Results and Discussion
The LPNTF + SRC pipeline significantly outperformed existing state-of-the-art methods. On the GTZAN dataset, it achieved a peak accuracy of 92.4%, a substantial improvement over the previous record of 82.5%. On the ISMIR 2004 dataset, it reached 94.38% accuracy.
Key insights from the results include:
- Structural Preservation: Maintaining the 3D tensor format during reduction prevents the loss of structural information inherent in flattened vectors.
- Manifold Learning: Incorporating local geometry via graph regularization is critical for the discriminative power of the features. LPNTF consistently outperforms standard NTF and other multilinear techniques across various levels of scatter retention.
- Robustness: The classification accuracy remained high even when the feature space was aggressively compressed to a fraction of its original size.
Conclusion
This research demonstrates that combining cortical-inspired features with manifold-preserving tensor factorization and sparse classification provides a powerful solution for music genre classification. Future work will focus on automated parameter optimization for the graph regularizer and evaluating the framework's performance in low-data regimes.
Summary updated