The Research On Chinese Keyword Spotting Technology For Real-Time Application

Keyword Spotting is an important area of speech recognition. Its objective is to spot the given keywords from continuous speech. Comparing with the continuous speech recognition, keyword spotting has advantages of less time consuming, higher calculation and stronger practicability and so on. Thus this technology has practical application in many areas.This paper is mainly focus on the Chinese keyword spotting for real-time application and achieving the keyword spotting system that has the good capability and can satisfy the real-time requirement.First, this paper generally introduces the framework and theory of keyword spotting. Secondly, we realize a keyword spotting baseline system. But the speed of recognition of this system can not satisfy the real-time requirement. Then, we improve the system for real-time application from two aspects. On the one hand, the restrictive model is presented. We provide the phone model restricting context called in-word- correlation phone model. This form of phone model can greatly reduce the size of model and increase the recognition speed. For reducing the high false alarm rate, we present the keyword verification based on acoustic confidence measure, and realize the multi-decoding mechanism. The good experiment result is obtained from the verification based on acoustic likelihood ratio. On the other hand, the decoding algorithm is improved. This paper analyses and researches the influence for decoding effect and speed by the beam pruning and maximum active model pruning. We investigate the best parameter that assure the good recognition capability and reduce the time consuming effectively. The verification mechanism is appended, and the experiment result is satisfying. Finally, in the conclusion, this paper discusses the future of the keyword spotting.