Detecting wake-word repeats in voice-activated systems

Abstract

In this paper we propose a practical method to improve the false rejection rate of a voice-activated system. The users of such systems tend to repeat the activation word in the case of the false rejection. In our method we modify a decoder of a phoneme-based voice-activated system for detecting such repeats which allows to catch 14.6% of previously undetected cases, while not increasing the false alarm rate, and 79% while adding less than 1% of false alarm rate. We use the fact that the repeat of the activation word consists of twice as many phonemes than the single pronunciation which allows more reliable detection.