Research On Speech Intelligent Synthesis Method Based On Spectrum Structure

Speech synthesis is an important research field in computer intelligent processing and human-computer interaction.At present,the data-driven method processes and synthesized speech signal by loading the smallest unit in the speech corpus.Although speech synthesis technology has a long research history,it is difficult to realize analog singing based on speech coding because of its low flexibility.Therefore,speech synthesis technology still has a lot of room for improvement.In order to study the frequency domain characteristics of the speech signal,fully excavate the speech spectrogram information,and realize the highly flexible speech coding technology,this thesis proposes a speech reconstruction method based on the frequency spectrum structure through the research on the human audiology principle and the research on skin-hearing aids priciple,and the effectiveness of the method is proved through experiments.This method mathematicalizes and formulates the sound data,so that the sound data can be stored and applied flexibly.The major research work of this thesis is as follows:(1)According to the principle of the skin-hearing aids,the speech signal of Chinese single vowel is analyzed.Combining the traditional short time analysis methods of speech signals and the principle of skin-hearing aids,the filtering effects of various filters are studied.This thesis carries out filtering experiments on Chinese single vowel speech signals and analyzes the frequency spectrum distribution of single vowel speech signals.The analysis shows that the pronunciation of Chinese phoneme[o]changes from[u]to[e],which is quite special,so when synthesizing phoneme[o],it must be combined with the frequency distribution and the mathematical expression function of phoneme[u]and phoneme[e].(2)A speech synthesis method based on spectrum construction is designed.The binary excitation model compulsorily divides the Chinese speech into unvoiced and voiced,which limits the diversity of synthesized signals in speech synthesis.Aiming at this problem,a speech synthesis method based on spectrum construction is proposed,which expresses speech signals in mathematical form and reconstructs speech signals through speech sinusoidal model,which can be used flexibly.Firstly,the voice signal is filtered to remove the low-frequency current signal interference;Secondly,the thesis uses Fourier Transform to transform the speech time domain signal into the frequency domain signal,that is,the speech spectrogram;Then,the thesis analyzes the frequency distribution of Chinese single vowel phonemes,and extracts the center frequency parameters and amplitude parameters;Finally,the WAVE file is used as the carrier of the voice signal.the WAVE file header is defined and speech synthesis functions are designed in C#language.Then,synthesis the speech signal.(3)An intelligent speech synthesis platform is built in C#language in the Visual Studio 2015 environment to synthesize speech signals.Firstly,starting from the functional requirements of speech synthesis by spectrum structurization method,the functional framework of the speech intelligent synthesis platform is designed;Secondly,initialize the file header parameters according to the WAVE file structure;Then,the voice signal parameters are loaded into the text box;Finally,the speech signal synthesis is realized by superimposing a single-frequency sine signal,and the WAVE file is used as the signal carrier,the file is written in a binary stream and the file is saved to the specified path.Comparing the synthesized speech signal with the original signal,the synthesized speech signal reduces redundant information,highlights the key frequency of the speech signal,and balances the energy of each frequency band.(4)Using synthetic speech signals to conduct subjective evaluation tests and establish speech function libraries.The subjective evaluation test uses the discrimination method to disrupt the sequence of the synthesized speech signal and the original speech signal.The testers performed the male and female voice discrimination test and the single vowel discrimination test.According to the subjective evaluation test results,the confusion matrix of synthesized speech is obtained.The test results show that the recognition rate of reconstructed Chinese phonemes[a],[e],[i],[u],[u]ranges from 83.3%to 88.9%and the recognition rate of Chinese phoneme[o]is 72.2%.The average recognition rate of Chinese single vowel phonemes is above 85%.Compared with the confusion matrix of Chinese single vowel speech by bispectral line reconstruction,the recognition accuracy rate of other phonemes is significantly improved except the phonetic factor[o].The frequency boundary of the Chinese single vowel phoneme simulation function is analyzed according to the frequency spectrum distribution of the synthesized speech signal and the confusion matrix,and a Chinese single vowel phoneme simulation function library is established.To sum up,this thesis design the spectrum structurization method to synthesizes speech according to the generation principle of speech signal,the principle of skin-hearing aids principal and the speech sinuaoidal model.The main frequency distribution of single vowel phoneme is obtained through experimental analysis,and the speech simulation function library is established.The article initially revealed the basic laws that speech signals can be represented by mathematical methods.The speech recognition experiment proves the feasibility of the spectrum construction method.The spectrum construction method has important theoretical significance and broad application prospects in speech intelligent processing and synthesis.