EffiAudioLiteNet: A Lightweight and Efficient Model for Audio Multi-Label Classification

Audio multi-label classification serves as a frontend application in audio processing, where the design of lightweight and efficient deep learning models becomes crucial for improved audio classification. Current architectures like Transformer carry excessive parameter counts, rendering them incompatible with resource-constrained devices such as mobile and edge devices. This study introduces an improved model named EffiAudioLiteNet, based on EfficientNet, which integrates a low parameter count with robust audio feature extraction, fitting well for audio classification tasks. For performance improvement, EffiAudioLiteNet replaces depthwise separable convolution with group convolution, integrates EMA attention, and swaps traditional MLP layers with the KAN structure, thus effectively enhancing model performance. This system achieves a 0.621 mean Average Precision (mAP) on the FSD50K dataset, with memory usage of just 8.1 MB, marking a $43 \% \mathrm{mAP}$ improvement over baseline methods on the same dataset.