Keyword Spotting with Quaternionic ResNet: Application to Spotting in Greek Manuscripts

Quaternionized versions of standard (real-valued) neural network layers have shown to lead to networks that are sparse and as effective as their real-valued counterparts. In this work, we explore their usefulness in the context of the Keyword Spotting task. Tests on a collection of manuscripts written in modern Greek show that the proposed quaternionic ResNet achieves excellent performance using only a small fraction of the memory footprint of its real-valued counterpart. Code is available at https://github.com/sfikas/quaternion-resnet-kws .