Showing 1–1 of 1 results for author: Morita, Y

Neural ODE with Temporal Convolution and Time Delay Neural Networks for Small-Footprint Keyword Spotting

Authors: Hiroshi Fuketa, Yukinori Morita

Abstract: In this paper, we propose neural network models based on the neural ordinary differential equation (NODE) for small-footprint keyword spotting (KWS). We present techniques to apply NODE to KWS that make it possible to adopt Batch Normalization to NODE-based network and to reduce the number of computations during inference. Finally, we show that the number of model parameters of the proposed model… ▽ More In this paper, we propose neural network models based on the neural ordinary differential equation (NODE) for small-footprint keyword spotting (KWS). We present techniques to apply NODE to KWS that make it possible to adopt Batch Normalization to NODE-based network and to reduce the number of computations during inference. Finally, we show that the number of model parameters of the proposed model is smaller by 68% than that of the conventional KWS model. △ Less

Submitted 6 September, 2020; v1 submitted 1 August, 2020; originally announced August 2020.

Comments: 5 pages, 5 figures