Sound event detection (SED) aims to classify target sound events from a given audio clip with the corresponding onset and offset time localization. SED, a core research field in machine listening, has been rapidly growing by adopting the methods by applying deep learning methods in image recognition fields. One representative example is 2D convolution which is applied to 2D time-frequency audio features. However, 2D audio data exhibits physical discrepancies from 2D image data, one of which is that frequency dimension of 2D audio data is translational-variant while position dimension of 2D image data is translational-invariant. Therefore, to make 2D convolution more consistent to 2D audio data, we need to make 2D convolution dependent to frequency dimension. Previously, to address this problem, Frequency Dynamic Convolution (FDYConv) was proposed to replaces 2D convolution. FDYConv is dependent on frequency dimensions since it applies different kernels on frequency dimension. While it has proven its outstanding performance on SED, there are several more alternative frequency dependent convolution methods which are yet to be explored, In this work, we propose other simpler frequency dependent convolution methods and compare them with FDYConv in order to further investigate the effect of frequency dependence of 2D convolution methods.
Read full abstract