The Research And Implementation Of Voice Conversion Technology

Voice is the most important tool in communication of human carrying abundant information. Voice conversion specifically change the personalized information without changing the source content. Voice conversion is a very important branch in the field of speech signal processing, The research of voice conversion has a high theoretical and application value.This paper conduct a intense research in voice conversion technology,and switch this technology into a system, it takes some researches and exploration work, build a multi-function voice processing and conversion system based on PC platform.Then, Expanding the software system into a DSP hardware platform with the TMS320C5509 A core, in this way, a voice conversion system based on DSP is completed. The main contents are as follows:(1) Elaborate the key technologies of voice conversion.Discuss the pre-filtering,endpoint detection methods,study how to use LPC root-finding method and short time autocorrelation function to the extract formant and pitch frequency parameters. Introduce the DTW algorithm and study how to complete a voice conversion with linear multivariateregression and average pitch conversion method. Also, voice synthesis algorithm and evaluation methods of the result is discussed in detail.(2) Build a multi-function voice processing and conversion system with Matlab GUI, M language and compiler tools. The system includes a real-time recording module, voice-loading module, pre-processing module, time domain analysis module,transform domain analysis module, voice conversion module,voice transformation module and other functional modules. The system can easily complete voice conversion with abundant functions and good flexibility.(3) After the completion of the software system, switch the system into DSP hardware platform and complete a DSP-based voice conversion system.(4) System test and experimental results show that: The system can achieve voice conversion well, from male to male,female to female, male to female,female to male,all are able to be converted correctly, and the result is good. Which turns out that similarity can reach to 86% in female-male experiment, and similarity also can reach to 81% in male-female experiments, the synthesized speech after conversion is clear and only has minimum distortion.