An End-to-End Far-Field Keyword Spotting System with Neural Beamforming

Abstract

Conventional keyword spotting (KWS) systems typically use microphone array techniques to improve robustness against noise and reverberation. However, KWS systems with such traditional speech enhancement methods may not always yield the optimal performance in different conditions, as the optimization criterion for the speech enhancement part is not directly relevant to the KWS objective. In this work, we explore the KWS system by encompassing neural beamforming for speech enhancement within the KWS neural network, which updates both parameters by jointly optimizing the unique KWS criteria. We demonstrate that the proposed neural beamforming KWS system not only significantly outperforms the traditional KWS method by improving 30% relative recall rate at the same precision, but also can obviously reduce the system complexity, which could be much easier to be adopted by small resource required devices.