Device Playback Augmentation with Echo Cancellation for Keyword Spotting

Keyword spotting (KWS) is required to operate in device playback conditions in which the device itself plays interfering signals. We propose a new method to augment the training set and adapt the acoustic model to the playback environment. It is based on acoustic simulation which models the coupling between the device’s loudspeakers and microphones. The employed model involves frequency response of the device, as well as room impulse response and nonlinear distortions introduced in the playback path. Finally, we pass the simulated signals through Acoustic Echo Cancellation (AEC) to model the artifacts introduced by AEC algorithm. The proposed method reduces False Rejection Rate in device playback noise by 25–60% for a Time-Delay Neural Network-based KWS engine. It is shown that the introduction of device characteristics and nonlinear filtration is necessary to achieve improvement in playback conditions. The augmentation scheme is highly independent of the architecture of the KWS system.