Identification and reconstruction of the unvoiced component in speech

Abstract

Speech is well described by a source-filter model. The source properties are critical for good quality reconstructed speech. We describe a source model which facilitates both low-rate coding and signal modification. The source signal is described by means of pitch-synchronous frame expansions, with different subsets of the coefficients corresponding to so-called voiced and unvoiced components. To obtain a perceptually plausible voiced-unvoiced decomposition even at speech onsets, our frame functions adapt to the signal. The generation of the unvoiced component consists of the replacement of the corresponding coefficients with realizations of a random variable with similar statistics. Existing sinusoidal and waveform-interpolation excitation models form approximations to the presented procedure.