Work

Voice personalization and speaker de-identification in speech processing systems

Year: 2019
Type: dissertation
Abstract: This PhD dissertation is part of the research proyect SpeechTech4All, dedicated to the advanced research in the main speech technologies in all the official languages spoken in Spain, the speaker's emotion recognition, and the construction of experimental frameworks to show the realized work. It is a coordinated proyect with the Polytechnic University of Catalunya and the University of the Basque Country. Specifically, the PhD dissertation falls within in the work package 1 (voice generation) of the proyect. At present, the group of Vigo has his own unit selection based TTS system in Galician and Spanish (called Cotovía and available as open source in http://sourceforge.net/projects/cotovia/). Throughout this proyect, experiments with HMM-based speech techniques has been started. Thus, the lingüistic information provided by Cotovía has already been integrated in the HTS system (http://hts.sp.nitech.ac.jp/), and the HMMs corresponding to the different speakers employed in Cotovía have been generated. The purpose of the dissertation is the development of new algorithms and/or improvement of existing algorithms in order to obtain a synthetic speech of higher quality and/or adapt the speech quality to the requirements of the application in which it will be used (a tradeoff among memory, computational load and quality). With this aim, we will work on unit selection and HMM-based synthesis techniques, and also with hybrid models (combination of unit selection and HMM-based techniques). In addition, in order to provide the speech synthesis systems with a larger number of speakers and speaking styles, we will work on speaker adaptation techniques. We also contemplate the possibility of using the speech synthesis systems to analyze the robustness of speech identification systems.
Language: English
Open Access status: green
OpenAlex ID: W2977136806
Cites: 218
Cited by: 1
Related to: 9