Keyword Spotting Based on CTC and Similarity Matching for Chinese Speech

Inspired by the end-to-end speech recognition systems and the deep learning technique, This paper presents a mandarin KWS system using an end-to-end approach, which can directly predict a posteriori of speech units. This system consists of two parts, the first part is a ASR model which is based on connectionist time classifier(CTC) and convolutional neural network and Long Short-Term Memory network. The second part is a keyword scorer which is based on lattice search space. The main difference between this KWS system and other KWS systems lies in the output letters which are syllables and corresponding keyword search mechanism. Through experiments, the results indicate that, 1) Character similarity matching method can improve the effect of keyword spotting;2) Long keywords have higher recognition accuracy than short keywords.