Age-Voice Conversion System Driven By Multi-Parameter

Voice Conversion is a technique used in order to change the personality characteristics of a source speaker’s voice into the target speaker’s and make the source speaker’s voice sound like the target speaker’s, while preserving the original semantic information. Age-voice Conversion which is one kind of Voice Conversion refers to change the age information of a source speaker’s voice into the target speaker’s and make the source speaker’s voice sound like the target speaker’s. This paper focus on the study of age-voice conversion based on speaker’s personality characteristics. Actually, we set up the system of age-voice conversion driven by speaker’s personality characteristics parameter. The main work and innovation of this thesis contains the following parts:(1)We established a little age voice database collected from a male and a female’s corpus recorded in different ages on the internet. Among them, the male’s corpus contains176sentences which were recorded at the male’s age of12,18and23. Also the female’s corpus contains85sentences which were recorded at the female’s age of12and20. The length of each sentence is between5seconds and10seconds.(2)We proposed an improved cross-pole linear prediction algorithm to solve the problem that formant extraction based on linear predictive coding (LPC) cannot handle the merge peaks and false peaks resulting in the less accurate of extracted formant frequencies. This algorithm can reduce the error caused by the cross-poles by modified the formant-pole’s radius, and then can enhance the accuracy of extracted formant frequencies.(3) As we all know, different voice track length is an important factor leading to sounding differences among the same speaker’s voice of different ages. To ensure the test voice can sound like the voice of target age after age-voice conversion, voice track length alignment is an indispensable technique. This paper analyzed two core issues that is frequency conversion factor estimation and frequency warping function selection in voice track length alignment technique, and constructed spectrum transformation model for the age-voice conversion. On this basis, we establish an age-voice conversion system driven by multi-parameter, and achieve a good effects about keeping personality characteristics of age-voice conversion.