Research On Method Of Unit Selection Speech Synthesis Based On Hidden Markov Model

Speech synthesis,which is the technology of text to speech(TTS),is an important branch of intelligent voice interaction.In today's rapidly development of information society,the ways of intelligent voice interaction has been widely pursued by people.Also,it has already been widely used in some application,such as the navigation of intelligent vehicle,the voice assistant of electronic equipment,bind reader and so on.These applications bring much convenience to people.There is no doubt that speech synthesis will play a very important role in people's life in the future.But at present the speech synthesis technology and expected goals still have a certain gap in the aspect of naturalness,which has seriously affected the further development of speech synthesis technology.Therefore,based on the research of speech synthesis,this paper will improve the traditional method and improve the naturalness of speech.At present the two popular methods of speech synthesis are speech synthesis based on statistical modeling and speech synthesis based on waveform splicing,they have their own advantage and disadvantage.The speech by the method of waveform splicing speech synthesis is more natural,which very close to the original sound.The method of based on statistical modeling synthesized speech can construct the system quickly,and has the stable speech.Besides,it needs only small storage space.Therefore this paper studies the above of two synthesis methods,combined with their advantage to study the method of unit selection speech synthesis based on Hidden Markov Model.In the aspect of until selection criteria,the traditional method used frame as the unit,which can easily lead to a decline in voice continuity.At the same time,the high complexity of the algorithm is inconvenient for practical application.Aiming at the deficiency,this paper will increase the selection unit.Using the sound finals as the unit to select.At the stage of splicing studied the PS OLA algorithm deeply,To improve the misjudgment of autocorrelation function method in the process of pitch estimation,using the center clipping function firstly,then using the autocorrelation function method.In order to simplify the calculation to improve the efficiency of the program,we can use the three level function instead of center clipping function.At the stage of unit splicing,the high frequency noise between the splicing point has a large impact on the naturalness of speech,so we add the corresponding transition element between the splicing unit to smooth the noise,the purpose is to improve the fluency and naturalness of speech synthesis.At the last of this paper making the comparison among the method of unit selection speech synthesis based on Hidden Markov Model and the speech synthesis based on waveform splicing as well as the speech of based on HMM parameters speech synthesis.The subjective evaluation and objective analysis of the three systems are compared,the results show that the speech naturalness of the this system is improved.