This paper proposes a new method for estimating the harmonic component of speech and speech-like signals. This class of signals is characterized by time-varying, wide-band spectra. Inherent in all calculations is therefore the assumption that the processing frame must be relatively small. Under this condition the Discrete Fourier Transform fails to provide accurate estimations. The method we propose consists of the parametric estimation of signal dynamics and a sophisticated spectral selection method which make the algorithm perform accurately with short frames. Moreover, contrary to existing Fourier methods, the harmonic estimates are not constrained to be multiples of the reciprocal of frame length, the resolution is therefore theoretically infinite. After the fundamental frequency has been estimated, the harmonic content can be easily reconstructed by fitting the estimates to the observations.
展开▼
