Accurate estimation of harmonic amplitudes in voiced speech based on harmonic transform

This paper presents a frame-based method for estimation of instantaneous amplitudes and frequencies of pitch harmonics. Proposed method assumes that the speech signal is a combination of voiced (harmonic) and unvoiced (noise) components present in whole speech band. Harmonic transform is used as a frequency analysis tool which kernel is synchronized with time-varying pitch frequency. The speech signal is divided into a set of subbands with center frequency corresponding to the pitch and in each subband Hilbert transform is used as an estimator of instantaneous amplitude of each harmonic.