Date of Award
Master of Science
Department of Electrical and Computer Engineering
Martin P. DeSimio, PhD
It is well known that there is room for improvement in the resultant quality of speech synthesizers in use today. This research focuses on the improvement of speech synthesis by analyzing various models for speech signals. An improvement in synthesis quality will benefit any system incorporating speech synthesis. Many synthesizers in use today use linear predictive coding (LPC) techniques and only use one set of vocal tract parameters per analysis frame or pitch period for pitch-synchronous synthesizers. This work is motivated by the two-phase analysis-synthesis model proposed by Krishnamurthy. In lieu of electroglottograph data for vocal tract model transition point determination, this work estimates this point directly from the speech signal. The work then evaluates the potential of the two-phase damped-exponential model for synthetic speech quality improvement. LPC and damped-exponential models are used for synthesis. Statistical analysis of data collected in a subjective listening test indicates a statistically significant improvement (at the 0.05 significance level) in quality using this two-phase damped-exponential model over single-phase LPC, single-phase damped-exponential and two-phase LPC for the speakers, sentences, and model orders used. This subjective test shows the potential for quality improvement of synthesized speech and supports the need for further research and testing.
DTIC Accession Number
Arb, H. Allan, "A Two-Phase Damped-Exponential Model for Speech Synthesis" (1996). Theses and Dissertations. 5919.