InterSpeech 2021

Acoustic Scene Classification using Kervolution-Based SubSpectralNet
(3 minutes introduction)

Ritika Nandi (MAHE, India), Shashank Shekhar (MAHE, India), Manjunath Mulimani (MAHE, India)
In this paper, a Kervolution-based SubSpectralNet model is proposed for Acoustic Scene Classification (ASC). SubSpectralNet is a competitive model which divides the mel spectrogram into horizontal slices termed as sub-spectrograms that are considered as input to the Convolutional Neural Network (CNN). In this work, the linear convolutional operation of SubSpectralNet is replaced with a non-linear operation using the kernel trick. This is also known as kervolution (kernel convolution)-based SubSpectralNet. The performance of the proposed methodology is evaluated on the DCASE (Detection and Classification of Acoustic Scenes and Events) 2018 development dataset. The proposed method achieves 73.52% and 75.76% accuracy with Polynomial and Gaussian Kernels respectively.