The Research Of Voice Conversion System Based On GMM With Independent Speaker Modeling

Voice conversion is changing one speaker's (source speaker) acoustic features to another (target speaker), then it was heard as if uttered by target speaker. It contains at most all the branches of speech signal processing area, so the research is useful for speech codec, speech synthesis, speech enhancement, speech recognition, etc. in this dissertation , an independent speaker modeling voice conversion algorithm based on GMM is proposed, the main work as follows:(1) proposed a new method based on independent speaker modeling, source and target acoustic features is modeled independent by GMM, this algorithm not only avoid the disadvantages of traditional method that needs parallel voice database, but also reduce complex degree of the system, especially in multi-people conversion system..(2) the system adopt dynamic rules in transformation process, which adjust rules real-time according to input speech data, so the spectrum conversion is more accurate than general method..(3) choose linear prediction synthesis platform, then use GMM training source and target lpcc features, at the end, transform the test speech, analysis the results and give the score.(4) choose STRAIGHT synthesis platform, then use STRAIGHT to get spectrum envelope, then use GMM training source and target features, transform the test speech, analysis the result and give the score.(5) use AUS theory to improve system performance, solute the problem that how to align two model components, It is very success after the experiment.(6) compared with traditional model., use the same wave to test independent model, then observed the results.