On Evaluating Formant Synthesizer Control Signals

Abstract

The problem of evaluating the inherent capability of a given synthesizer to produce good speech is inseparable from the problem of evaluating the signals that control it. Those values of control signals that give the best speech output must be accepted as optimum, and the resulting performance of the synthesizer must be accepted as maximum. Optimizing the control signals must be done, however, not in some average fashion, but with reference to the specific phonemic content of the speech. The objectives of psychoacoustic test procedures for this purpose are thus somewhat different from those used in preference testing of speech communication systems, and so are the means. Forced-choice paired comparisons have been used to evaluate different control signals for one synthesizer producing specific phonemes in specific environments. The results are used as a guide for choosing the best means for deriving the signals from real speech. Evaluations of the control signals for stop consonants from a Glace-Holmes Synthesizer will be presented.