Abstract
Traditional music genre classification systems often rely on "flattening" multi-dimensional audio features into one-dimensional vectors, a process that discards inherent structural correlations and exacerbates the curse of dimensionality. This paper introduces a multilinear approach that treats audio features as third-order tensors. By applying multilinear subspace techniques—specifically Non-Negative Tensor Factorization (NTF), High-Order Singular Value Decomposition (HOSVD), and Multilinear Principal Component Analysis (MPCA)—to bio-inspired spectro-temporal features, the authors demonstrate that preserving the data's multi-dimensional structure significantly enhances classification performance.
Introduction
Music genre classification is a fundamental task in Music Information Retrieval (MIR), yet it remains challenging due to the subjective and often inconsistent nature of genre labels. A primary technical hurdle is feature representation. Standard systems extract features from short audio segments and concatenate them into high-dimensional 1D vectors. For instance, a representation with 6 scales, 10 rates, and 128 frequency bands results in a 7,680-dimensional vector. This "vectorization" leads to two major issues: the loss of natural correlations between time, frequency, and modulation, and the "curse of dimensionality," where high-dimensional spaces require impractical amounts of training data to avoid overfitting.
To address these challenges, the authors propose a multilinear framework that represents music as a 3rd-order tensor (a "data cube"), thereby maintaining the relationships between different acoustic dimensions throughout the analysis process.
Methodology
The proposed pipeline integrates bio-inspired feature extraction with tensor-based dimensionality reduction.
Bio-inspired Feature Extraction
Instead of standard Mel-frequency cepstral coefficients (MFCCs), the researchers utilize a computational model of the human auditory system. This model generates a cortical representation capturing:
- Scales: Spectral modulations (how sound changes across frequencies).
- Rates: Temporal modulations (how sound changes over time).
- Frequencies: Standard spectral content.
Averaging these over time results in a $6 \times 10 \times 128$ tensor for each audio track.
Multilinear Subspace Analysis
Three multilinear techniques were evaluated for dimensionality reduction without flattening the tensor structure:
- Non-Negative Tensor Factorization (NTF): Decomposes the tensor into rank-1 components, facilitating "parts-based" representations.
- High-Order Singular Value Decomposition (HOSVD): A tensor generalization of PCA that identifies principal axes for each dimension independently.
- Multilinear Principal Component Analysis (MPCA): A multilinear equivalent of PCA designed to capture maximum variation across the tensor ensemble.
Following dimensionality reduction, the compact features are classified using a Support Vector Machine (SVM) with a Radial Basis Function (RBF) kernel.
Experimental Results
The approach was evaluated on two benchmark datasets: GTZAN (1,000 recordings, 10 genres) and ISMIR 2004 Genre (1,458 recordings, 6 genres).
The results indicate that NTF and HOSVD are particularly effective at capturing the nuances required for genre classification. On the GTZAN dataset, NTF achieved an accuracy of 78.20%, while HOSVD reached 80.95% on the ISMIR 2004 dataset. These results are highly competitive with contemporary state-of-the-art methods.
As shown in the comparison table, the multilinear approach outperforms several established benchmarks, such as Tzanetakis et al. (61%) and Lidy et al. (76.8%). While Bergstra et al. reported a slightly higher 82.5% on GTZAN, the authors highlight that their multilinear method is unsupervised during feature extraction, potentially offering better generalizability across different musical domains.
Discussion and Conclusion
The primary insight of this work is that dimensionality reduction is most effective when it respects the inherent tensor structure of the data. By operating on the modes of the tensor separately, the physical meaning of features—such as frequency versus temporal rate—is preserved. The superiority of NTF and HOSVD over MPCA suggests that finding localized patterns or orthogonal subspaces is more critical for genre identification than simply capturing total variation.
While the results are promising, the authors note that the current tensor decomposition methods are unsupervised. Future work could explore supervised multilinear techniques to further enhance accuracy. Additionally, addressing the "small sample" problem in multilinear algebra and exploring multi-label classification for ambiguous genres remain important avenues for research.
Summary updated