Improved Keyword Spotting Based On Keyword/Garbage Models

With the development of computer hardware and software technologies,we have gradually entered the era of artificial intelligence.Thanks to mobile Internet and the Internet of Things,all beings start to interoperate.Human-computer interaction has become the focus technological developments.Voice interaction is an important form for human-computer intelligent interaction.Voice is the most natural and convenient communication method for human beings.Thus voice-based human-machine intelligent interaction plays an important role in the current development of science and technology.The ultimate goal of voice interaction is to achieve barrier-free communication between people and machines.With decades of technology development in the field of speech recognition,this goal is being realized.A large number of effective algorithms have laid a good foundation for voice interaction.Based on hidden Markov model,Gaussian mixture model and token passing algorithm,this thesis develops a speech keyword spotting system for both Windows and Android platforms.Based on the uncertainty of the decoder we found in our experiments,optimization methods are proposed to greatly reduce the probability of false alarm,while maintaining the system's efficient and accurate recognition of predefined keywords,enabling the system to be used in smart homes for fully hands-free control of televisions or other smart devices.The related work and innovation of this article are as follows.Based on the existing speech acoustic modeling of Hidden Markov Model,a Windows platform keyword recognition system was developed.We study the drawbacks of the token passing decoding algorithm and introduce the keyword/garbage models into the system in order to to substantially reduce false alarm rates.Through the detailed analysis of the experimental results,we also propose two confidence assessing methods for degree of match and degree of stability to further optimize the system's performance.Our experimental results have verified the effectiveness of the proposed methods.Our experiments show that these two methods can substantially reduce the false alarm rate from 75.05% to 5.71%.Meanwhile,the recognition accuracy of keywords is almost unchanged.Finally,after improving the system performance,we also transplanted the speech recognition system of Windows platform to Android through JNI interface and develops the user interaction interface in order to increase the application scenarios of the system.