Abstract
This paper presents a novel Convolutional Neural Network (CNN) architecture designed for efficient and accurate music genre classification. By employing "duplicated" convolutional layers that feed into parallel max-pooling and average-pooling paths, the proposed "Net1" model preserves a richer set of statistical features from mel-scale spectrograms compared to standard residual networks. Evaluated on the GTZAN dataset using 3-second audio segments, the model achieves a classification accuracy of 90.7%, outperforming several established benchmarks and demonstrating particular strength in identifying genres with distinct timbral characteristics.
Introduction
Classifying music by genre is a foundational task for digital library organization and recommendation engines. Historically, this task relied on "hand-crafted" features—expert-designed mathematical descriptions of rhythm and timbre—which often fail to generalize across diverse musical styles. While deep learning has automated feature extraction, standard CNN architectures can lose critical information during pooling or addition operations, and complex models like Convolutional Recurrent Neural Networks (CRNNs) are often computationally expensive.
This research addresses these challenges by optimizing how information is processed and retained through the network. The authors propose a branching structure that utilizes duplicated layers to capture both prominent features and general statistical distributions simultaneously.
Methodology
The core innovation of this study is the Net1 architecture, which departs from linear layer stacking in favor of a branching design.
Input Representation
The model processes audio as mel-scale spectrograms (256x128). Unlike standard Hertz-scale representations, the mel-scale mimics human pitch perception, emphasizing psychologically relevant frequency distributions that are crucial for genre identification.
Duplicated Convolutional Layers
After initial convolutions, the network duplicates the output into two parallel paths. One branch undergoes max-pooling to identify the most prominent acoustic features, while the other undergoes average-pooling to capture the broader statistical distribution of the signal.
Feature Concatenation
Inspired by, but distinct from, Residual Learning (ResNets), Net1 concatenates the outputs of these parallel pooling layers rather than adding them. This approach increases the dimensionality of the data fed into the final dense layers, ensuring that distinct statistical features are preserved for the final classification stage.
Results and Discussion
The study utilized the GTZAN dataset, consisting of 1,000 song excerpts across 10 genres. The 30-second clips were processed into 3-second segments to test the model's efficiency with short-duration audio.
Performance Metrics
The proposed Net1 model achieved a top-tier accuracy of 90.7%, significantly outperforming a baseline residual-addition model (Net2, 88.3%) and other established methods like KCNNs (83.9%). Notably, Net1 achieved this with fewer parameters (~1M) than many residual-based counterparts (~1.25M), suggesting that the connectivity strategy is more vital than raw model depth.
Genre-Specific Analysis
As shown in the confusion matrix, the model achieved 100% accuracy in identifying Classical, Jazz, and Metal. The primary challenge remains the "Country" genre, which was frequently misclassified as Rock (21% of cases). This is likely due to the 3-second window being insufficient to capture the long-term rhythmic structures that distinguish Country from similar styles.
Conclusion
The use of duplicated convolutional layers and parallel pooling concatenation provides a robust framework for music feature extraction. Future work will explore end-to-end learning directly from raw audio waveforms and the integration of Recurrent Neural Networks (RNNs) to better capture the temporal dependencies necessary for improving classification in rhythmically complex genres like Country music.
Summary updated