Low-complex and Highly-performed Binary Residual Neural Network for Small-footprint Keyword Spotting

The hardware power-aware Keyword Spotting (KWS) implementation requires small memory footprint, low-complex computation, and high accuracy performances. In this article, three aspects are introduced to satisfy these three stringent requirements. Firstly, a lightweight Binary Residual Neural Network (B-ResNet) is proposed and applied to the small-footprint KWS. The parameters and calculations inside the network are greatly downscaled during the binary quantization. Secondly, during the forward propagation, distribution of the binary activation is optimized by our proposed learnable activation function with fix-valued shift initialization. Thirdly, our variable periodic window (PW) for the backward gradient correction (BGC) is also put forward to avoid gradient mismatch and vanishing problems during the back-propagation. These two improvements effectively increase the accuracy performance during the binarization. Our studies in this article are very helpful and promising for the future hardware KWS implementations.