Voice Conversion Based On Isolated Speaker Model

Voice conversion technique attempts to modify the source speaker's voice to sound as if it was uttered by the target speaker. Voice includes semantic information and the speaker's individual information. In voice conversion the acoustic parameters related to speaker's individuality are transformed while maintaining original semantic information, making the newly synthesized voice is closer to the target speaker. Conventional voice conversion systems are usually based on parallel speech corpus and joint training and then derive the appropriate voice conversion rules, but it is difficult to get parallel speech data and inflexible to extend system in application. In this dissertation, voice conversion based on isolated speaker model(ISM) is proposed, the main work as follows:1. Research STRAIGHT analysis and synthesis platform, then use STRAIGHT to modify the acoustic parameters related to speaker's individuality.2. Construct conventional GMM based voice conversion system with parallel corpus and analyze the characteristic of GMM.3. Proposed structured Gaussian mixture model(SGMM) to describe the characteristics of the speaker's acoustic space.4. Use Acoustical Universal Structure (AUS) to align the Gaussian distributing of source and target speakers'SGMMs.5. Given the framework of voice conversion based on ISM framework, and programming the entire system.The ABX and MOS experiments show that the proposed method has the equivalent conversion performance as conventional method, and the speaker recognition rate of transformed voice reach to 94.5%. All in all, the method proposed in this dissertation not only has a good conversion performance, but also has a smaller amount of training and good system extensibility.