Adaptive Noise Suppression for Wake-Word Detection by Temporal-Difference Generalized Eigenvalue Beamformer

This paper proposes an adaptive noise suppression method for wake-word detection by a temporal-difference generalized eigenvalue (TDGEV) beam-former. To emphasize wake-word utterances, which are leading phrases with short duration, the proposed method is based on a generalized eigenvalue beamformer regarding current and past spatial covariance matrices for speech and noise, respectively. It can emphasize wake-words with small distortion and suppress any noises regardless of directions of arrival (DoAs) and noise sources. We perform experiments of wake-word detection with and without beamformers using test data including wake-word utterances from various DoAs. The results show that the proposed TDGEV method reduces false rejects with 32.9% relative error rate reduction.