Research On Continuous Speech Recognition Technology Based On HMM

In the 21st century,artificial intelligence developed rapidly,and speech recognition technology was an important aspect of artificial intelligence.With the development of hardware and software technology,the technology of continuous speech recognition has made great progress.The performance of a continuous speech recognition system is related to two important factors,one is the speech recognition model,and the another is the segmentation of continuous speech.For speech recognition model,we can choose Hidden Markov Model with strong processing ability for temporal signals or Artificial Neural Network with autonomous learning ability.After analyzing three commonly used speech recognition models,this paper selects the Hidden Markov Model to study the Chinese continuous speech recognition.The segmentation of continuous speech is a great difficulty in speech recognition technology.Under the condition of massive training speech,the segmentation based on model can achieve continuous speech segmentation to a certain degree.There are a lot of problems when training voice inadequacies.This paper analyzes the pronunciation characteristics and the voice structure of Chinese,and then the multilevel segmentation method of Chinese continuous speech is studied by using the language spectrum and the pitch cycle trajectory.The specific research contents of this paper are as follows:(1)Feature analysis of speech signal.The segmentation is a difficult point for Chinese continuous speech recognition.In order to realize the segmentation of continuous speech,first of all,we need to understand the characteristics of the Chinese speech signal.This paper analyzes the characteristics of Chinese speech signals in different domains,analyzes the spectrum characteristics of speech signals by spectrogram,and draws the pitch period trajectory of voiced signals by cepstrum features.(2)Contrastive analysis of speech recognition models.In this paper,three speech recognition models are studied,namely Vector Quantization,Gauss Mixture Model and Hidden Markov Model.These recognition models are tested by existing voice files,and the advantages and disadvantages of each model are analyzed.(3)A multilevel segmentation method for Chinese continuous speech.In this paper,the endpoint detection technology based on the time domain characteristic parameters and the technique of endpoint detection based on cepstrum are studied,and analyze the shortcomings of the above speech endpoint detection technology firstly.Then,based on the analysis of the characteristics of Chinese speech signals,this artical has studied the multilevel segmentation method of Chinese continuous speech by using coherence analysis,pitch cycle trajectory detection,and gray-scale mean analysis of spectrogram.In this paper,the segmentation of Chinese speech under the microphone channel is made by multilevel segmentation,and the accuracy of segmentation is about 91%.Compared with the segmentation method based on time domain characteristics and the segmentation method based on frequency domain characteristics,the effect of multilevel segmentation method is greatly improved.