Transform Method of Fast Digital Synthesis of Unvoiced Speech Sounds

Abstract

Digital speech synthesis is often carried out by direct convolution of a sequence of unit samples with a sampled impulse response. In the case of unvoiced sounds, this method can lead to the need for an excessively large number of additions per output sample of the synthesized waveform. This is because of the requirement that the synthesized output be essentially Gaussian and possess the correct spectral energy distribution in order to achieve natural-sounding unvoiced sounds. This paper describes a transform method based upon the Cramer-Kolmogarov representation for a stationary process that reduces the required number of additions per output sample by a factor of at least 10, and which has been found experimentally to produce high-quality synthetic speech. The performance of the method is demonstrated by playing a tape recording of the output of an all-digital vocoder that uses this method of synthesis. [Work supported by Air Force Avionics Laboratory, Wright-Patterson Air Force Base, Ohio.]