Two-stage Strategy for Small-footprint Wake-up-word Speech Recognition System

In this paper, we propose a small-footprint wake- up-word speech recognition (WUWSR) system with two stages to recognize a two-syllable wake-up word. In the first stage, convolution neural network (CNN) is trained to predict the posterior probability of context-dependent state. Thus a wake-up-word is detected according to the confidence score obtained by dynamic programming. In the second stage, we cascade bidirectional long short-term memory network (LSTM), convolutional modules and deep feed-forward network (BLCDNN) successively to verify the detection. The first stage quickly filters out speech without wake-up word, and the second stage refines the detection. In addition, without the intervention of any decoding modules, the proposed system can guarantee low latency. The experimental results demonstrate the effectiveness of this method. Our system, named CNN-BLCDNN, reaches high accuracy and maintains low false alarm rate.