Speech morphing by independent interpolation of a spectral envelope and source excitation

Abstract

A speech morphing algorithm based on progressive interpolation of spectral envelopes and source signals is proposed. The basic morphing scheme is; 1) determine the time correspondence for unit waveforms of original and target speech, 2) separate speech spectra into an envelope and a sound source, 3) obtain the frequency correspondence for spectral channels of original and target speech for each envelope, 4) interpolate both source signal and envelope, 5) construct a unit waveform, and 6) generate morphing speech by PSOLA. In the objective test, the proposed method can reduce the spectral distortion by 1.9 dB in comparison with the method based on progressive substitution of spectra, when it is used for interpolating two vowels in a real speech. The effectiveness of the method is also confirmed by a subjective test in which 89% (male to female) or 93% (female to male) subjects preferred the proposed method. © 1998 Scripta Technica, Electron Comm Jpn Pt 3, 82(3): 22–30, 1999