Audio classification with Dilated Convolution with Learnable Spacings

TL;DR

This paper demonstrates that dilated convolution with learnable spacings (DCLS) enhances audio classification performance across multiple architectures on the AudioSet benchmark without increasing model complexity.

Contribution

It introduces the application of DCLS to audio tagging, showing significant improvements in accuracy for existing architectures without additional parameters.

Findings

Improved mean average precision (mAP) on AudioSet

No increase in model parameters

Low throughput cost

Abstract

Dilated convolution with learnable spacings (DCLS) is a recent convolution method in which the positions of the kernel elements are learned throughout training by backpropagation. Its interest has recently been demonstrated in computer vision (ImageNet classification and downstream tasks). Here we show that DCLS is also useful for audio tagging using the AudioSet classification benchmark. We took two state-of-the-art convolutional architectures using depthwise separable convolutions (DSC), ConvNeXt and ConvFormer, and a hybrid one using attention in addition, FastViT, and drop-in replaced all the DSC layers by DCLS ones. This significantly improved the mean average precision (mAP) with the three architectures without increasing the number of parameters and with only a low cost on the throughput. The method code is based on PyTorch and is available at https://github.com/K-H-Ismail/DCLS-Audio