Temporal Blueprint Separable Convolution for Efficient Keyword Spotting

Abstract

Keyword Spotting (KWS) is a task that detects wake-up words or distinguishes commands in a stream of audio. Since this type of task is normally used on low-resource devices, such a task requires an implementation that has a small memory footprint and low power usage. In this thesis, temporal blueprint separable convolutions are presented as highly efficient blocks that can be incorporated into Convolutional Neural Networks (CNNs) used for KWS. Based on intra-kernel correlation properties from trained CNN models, temporal blueprint separable convolutions are shown to more efficiently separate regular temporal convolutions than temporal depthwise separable convolutions. In addition to temporal blueprint separable convolutions, group convolutions with channel shuffle and a Multi-Scale Temporal Blueprint (MTB) method is used in the proposed approach. It is experimentally shown that the proposed KWS models achieve high accuracy with a small footprint.