Voice conversion based on static speaker characteristics

• DOI: 10.1109/comsig.1998.736922
• Publication Date: 2002-11-27T18:21:56Z
• Authors: L.C. Schwardt; J.A. Du Preez
• Keywords: Speaker identity; Statistics; Speech recognition; Dynamic programming; 01 natural sciences; 410; Speech synthesis; 400; Pitch extraction; Speech processing; 0103 physical sciences; Speech analysis; Algorithms; Voice conversion

Voice conversion has recently emerged as an interesting branch of speech processing that deals with the modification of a speaker's perceived identity. This technology has applications in speech recognition, the entertainment and security industries. This paper provides a brief introduction to current voice conversion approaches, and discusses the development of the PASS system, a parametric voice conversion algorithm based on static speaker characteristics. The system is easy to implement, requires no phonetic transcription of the speech data, and is shown to be valuable in the case where very little training data is available. Particular mention is made of the pitch extraction subsystem, which uses a novel pitch determination algorithm to ensure the robust estimation of pitch statistics.