New Methods for Voice Conversion

iii ACKNOWLEDGEMENTS I would like to thank to my thesis supervisor Assoc. Prof. Levent M. Arslan for his guidance and support. It was a great pleasure to work with him during this thesis and in all the projects we have been involved in the last three years. I would like to thank Prof. reading my thesis and participating in my thesis committee. I would like to thank to my family and to Aylin for their endless love, encouragement, and support. I would like to express my gratitude to Barış Bozkurt for his inspiring ideas. Special thanks go to my colleagues at Sestek Inc. And to everyone who has participated in the subjective tests: Thanks for your patience and time. I promise, the tests will take shorter next time. iv ABSTRACT NEW METHODS FOR VOICE CONVERSION This study focuses on various aspects of voice conversion and investigates new methods for implementing robust voice conversion systems that provide high quality output. The relevance of several spectral and temporal characteristics for perception of speaker identity is investigated using subjective tests. These characteristics include the subband based spectral content, vocal tract, pitch, duration, and energy. Two new methods based on Wavelet Transform and selective preemphasis are described for transformation of the vocal tract spectrum. A new speaker specific intonational model is developed and evaluated both in terms of accuracy and voice conversion performance. A voice conversion database in Turkish is collected and employed for the evaluation of the new methods.