Study And Implementation Of Speech Modification

Speech modification is the course that a speaker's voice is modified to sound like another person. According to scientists's research, among the characteristic parameters of speech, pitch period is the most important one to contribute for the identity of speech while formant frequency is the second important one. This paper's work completes a speech modification system which can change speech's pitch period and formant frequency.The paper's main works are as follow:(1) Several methods of extracting pitch period are researched. The autocorrelation algorithm is adopted to extract pitch period among these methods. This method confirms each speech frame's pitch period candidate by searching the max autocorrelation and then finds the most appropriate pitch period by using Viterbi algorithm.(2) Study the method of labeling pitch impulses. Find the speech's voiced part's point at which there is the max peak, then this point will be regarded as a basic point, we search other pitch impulses which have the max autocorrelation value from its left side and right side.(3) The Time Domain Pitch Synchronous Overlap Add(TD-PSOLA) method is researched. After pitch period's modified, the speech signals will be reconstructed according to the pitch impulses. The Hannig window is used to reconstruct speech and the time scale of speech is also modified according to the formant frequency's changing factor.(4) Resample the speech which has been reconstructed. The speech's resampling is based on the method that any point of a speech in time domain can be reconstructed through the interpolation of the original speech, it changes the number of sampling points, and the speech will be played in original sampling rate. Therefore, the speech is stretched or compressed in time domain througth resampling, which changes the pitch period and formant frequency of speech.(5) Complete a whole speech modification system wihch can adjust the original speech.At present there are two main problems for speech modification: The first one is that the reconstructed speech's naturalness will decline when the speech parameter is changed greatly. The second one is that the modification of formant frequency caused by speech resampling can't be contolled well. These problems need further research, with the development of teconology, better speech modification system will appear.