Improved Spectral Control in Rule-Based Hybrid Text-To-Speech Synthesis for the Production of Natural-Sounding and Mimetic Voices

This project demonstrates the feasibility of a novel text-to-speech synthesis system that provides individuals with speech impairments with a selection of more natural-sounding and individualized voices. Through a spectral envelope (SE)-based synthesis technique to be integrated with NovaSpeech’s hybrid synthesis system the project provides improved control over the spectral cues to voice quality. The representation of syllable nuclei in terms of spectral envelopes rather than the waveforms currently used in the hybrid system provides for greater adaptability to different linguistic contexts and speaking styles. The SE-based representation promises improved naturalness and mimesis while maintaining all other benefits of the hybrid system, including rapid and low-cost voice development, consistent quality from utterance to utterance, and small memory requirements. Since the system does not trade off natural-sounding speech for low memory requirements, as current systems do, it is able to meet the demands of virtually any text-to-speech application, such as a talking cell phone or e-reader, a screen reader for individuals with visual impairments, or a larger server-based system providing information over the phone. The ease with which natural-sounding and individualized voices can be added with a combined spectral envelope and hybrid approach is of special benefit to individuals with speech impairments. The project also has the advantage of increasing understanding of speech perception, speech patterns, and speaker identification.