MRNet: A multi-route convolutional neural network for robust music representation learning

Music Information Retrieval (MIR) focuses on extracting semantic information embedded in audio signals, such as genre, artist identity, and tempo. These musical cues cover a wide range of temporal characteristics, from short-term features like pitch and timbre to long-term patterns such as melody and mood, and they require processing at multiple levels of abstraction. In this paper, we propose a Multi-Route Neural Network (MRNet) designed to capture musical representations that reflect both short-term and long-term characteristics, as well as different levels of abstraction. To achieve this, MRNet stacks several convolutional layers with different dilation rates, allowing the model to analyze audio patterns over multiple time scales. Additionally, it introduces a specialized module called the multi-route Res2Block, which separates the processing path into multiple branches. Each branch processes the input to a different depth, enabling the network to extract low-level, mid-level, and high-level features simultaneously. MRNet achieves classification accuracies of 94.5 %, 56.6 %, 63.2 %, and 71.3 % on the GTZAN, FMA Small, FMA Large, and Melon datasets, respectively, outperforming previous Convolution Neural Network(CNN)-based approaches. These results demonstrate the effectiveness of MRNet in learning robust and hierarchical music representations for MIR tasks.