Automatic generation of voiceless excitation in a vocal cord-vocal tract speech synthesizer

Abstract

A speech synthesis technique is described which incorporates acoustic models for sound propagation in a tube with yielding walls, turbulent noise generation at locations of constricted volume flow in the vocal tract, and the self-oscillatory properties of the vocal cord source. This formulation frees the experimenter from a traditional limitation, namely, the assumption of linear separability of sound source and resonant system. As a consequence, new opportunities accrue for building realistic physiological characteristics into the synthesizer. These built-in characteristics represent information that need not be overtly supplied to control the synthesizer. The system is used to synthesize test syllables from controls which are stylized models of articulation and connected speech from controls automatically derived from printed text. The synthesis technique demonstrates the feasibility of generating all speech sounds (voiced, unvoiced, nasal) from a common set of physiologically based control parameters, as the human does.