Design And Implementation Of Offline Speech Recognition Algorithm

In recent years,with the popularity of phonetic entry method,voice assistant,intelligent speaker and other emerging products,speech recognition technology has gradually entered people's vision.While most of the existing speech recognition platforms are complex in structure,high in power consumption,high in cost and being dependent on the cloud voice library.So that they are difficult to be applied in more cases.Local speech recognition,by discarding its universality and optimizing the algorithm for the special offline application scenarios,has decreased the threshold of speech recognition technology application.It has far-reaching research significance and wide application space.This paper focuses on the research and design of a speech recognition algorithm based on Hidden Markov chain model applying in the special offline embedded systems.First of all,the generation mechanism and characteristics of the voice signal is analyzed,which provides the theoretical basis for the later establishment of the speech recognition model.Then we analyze the related technologies of speech signal preprocessing,and design a variable step size block frequency-domain LMS filter noise reduction algorithm and a double threshold endpoint detection algorithm based on the short-time average energy and short-time average cross zero ratio.Then we propose a HMM model for speech recognition modeling,solve the three classic problems in the application of HMM model,and design a new training algorithm for getting right model parameters.Finally,we design a set of speech recognition scheme applying in offline scenario based on HMM model.The speech recognition system is realized on FPGA So C platform,which can recognize simple speech instruction in real time based on the research of local speech recognition algorithm.It is able to train the parameters of HMM model in real time.Tests' results show that the accuracy rate of recognizing the simple Chinese speech instruction phrases is more than 95%.This will lay a solid foundation for further research on speech recognition technology and development of an offline speech recognition chip.