We propose to develop a system for representing speech using time-varying control parameters for a source-filter model of speech synthesis. We would employ techniques of acoustic analysis in a known phonetic context to extract KLSYN88 synthesizer parameters from speech waveforms. Such a synthesis-based encoding would emphasize meaningful acoustic-phonetic properties of speech, making it possible to selectively modify specific production-related attributes, and allowing immediate resynthesis of the original or modified waveforms. Phase I has two goals. First, to determine how well KLSYN88 parameters can represent specific natural utterances. Second, to determine whether suitable parameter values can be extracted from natural speech using acoustic algorithms operating in known phonetic contexts. If successful the resulting technology would be employed as a speech training aid for normal-hearing subjects with speech impairments, allowing them to hear corrected versions of their own utterances. It would also be integrated into a commercially-supported version of KLSYN88, thus substantially reducing the amount of time required to generate control parameters.
