Viterbi Decoding Based On Token Passing And Its Application In Speech Recognition

Nowadays, the technology of mobile internet develops rapidly, the use of smart terminals deeps into every aspect of human daily life day by day. The direct means of communication between human and smart terminals is speech communication. Therefore, the technology of speech recognition has become an important symbol of the modern science and technology. Meanwhile, it has also become one of the important area of the research and development in the mobile internet. The ultimate goal of speech recognition is to realize the barrier-free direct communication between human and machine. The theory of speech is becoming more and more mature after half a century’s development. A large amount of practical applications have indicated that many speech recognition models and decoding algorithms are pretty helpful by far.This article, starting from the conventional models of speech recognition and decoding algorithm based on the Hidden Markov Model and token passing decoding algorithm, describes the development of a speech recognizer working on the Windows and Android platform. This decoder can be applied into mobile devices, home automation, wearable devices and vehicle equipment. The detailed work and innovation points of this research are as follows.Firstly, we introduce the technology of speech recognition, describe the background of the selected topics, research significance, and the overseas and domestic research status of this area. Then, we discuss Hidden Markov Model, expectation maximization algorithm, and the three basic problems of Hidden Markov Model. After that, we focus on the design and realization of speech recognition which works on the Windows platform. We describe five modules of this system with the combination of the whole frame system structure, starting from the development tools and database system. We then give a detailed introduction of the construction of the real-time speech recognition system, and the method and process of system transplantation. The performance and quality of this system indicate that not only both recognition rate and real time factor of this system can achieve application requirements, but also both the consumption of the system resource remains low and the practicability of this system is strong. While in the noisy environment, the robustness of our proposed system outperforms the mainstream speech recognition products appear in the market. Finally, to overcome the shortage of the token passing algorithm, we propose optimization methods: 1) the pruning algorithm; 2) limit the maximum number of active tokens.