Involution Based Speech Autoencoder: Investigating the Advanced Vision Operator Performance on Speech Feature Extraction

Feature extraction is an important aspect of deep learning. In recent studies, various methods from image processing have been applied to speech feature extraction. Although the involution operator has achieved great success in vision recognition, the use of the involution operator in speech tasks has not been investigated. To determine whether the performance improvement of this operator on visual tasks is reflected directly in the speech processing domain, this study investigated autoencoder networks operating on Mel-spectrograms utilizing involution and convolution. It was determined that convolution continues to outperform involution as a direct replacement. The reasons for discrepancies in the input feature dimensions of different fields are discussed. Based on its channel-agnostic property, involution in combination with convolution requires additional research.