Keyword spotting based on recurrent neural network
Zhou jianlai, Jian Liu, Song Yantao, Tiecheng Yu · 2002
In this paper, a recognition method is proposed which has been applied to KWS (keyword spotting) and based on a recurrent neural network. The proposed recurrent neural network has memory within a several time steps' interval, and its classification ability is more powerful than the HMM (hidden Markov models). It can process a time sequence signal and shows some interesting properties in performing time warping. The RTRL (real time recurrent learning) algorithm is used to train the neural network. Two numerical examples are presented to demonstrate the merit of the neural network. Finally, we look at the defects of the scheme, and discuss strategies for combining this approach with HMM in order to decide the time position of KWS by in a more precise manner.