Efficient representation of short-time phase based on time-domain smoothed group delay

A method for representing the short-time phase of speech efficiently is proposed. In this paper, we attempt to represent the short-time phase using a time-domain smoothed, group delay (TSGD) obtained by scaling the inverse Fourier transformed speech group delay by an appropriate order. The results of tests using segmental SNRs show that information regarding the shapes of waveforms is concentrated at TSGDs of orders less than 100. In addition, subjective evaluation tests show that high-quality speech almost indistinguishable from the original speech is synthesized when the order of the TSGD is about 30 and the short-time phase can be represented efficiently by the proposed scheme. In addition, since the proposed scheme also suppresses degradation due to errors associated with speech-no speech decisions, which has been a problem in past analysis-synthesis schemes, it is considered to be useful in applications such as speech coding and speech quality transformation. (C) 2003 Wiley Periodicals, Inc.