Automatic and reliable estimation of glottal closure instant and period

The authors present an automatic and reliable algorithm for determining glottal closure instant (GCI). As a byproduct, nonstationary fundamental period estimation is achieved. The computation includes twelve-pole speech linear-prediction analysis, cross correlation, and both direct and inverse fast Fourier transforms (or a convolution). Maximum-likelihood epoch determination is used as the basis for locating GCIs, and the Hilbert transformation is applied to improve performance and reliability. A description of the system and the voiced/unvoiced/mixed (V/UV/M) decision procedure is given. The results show that the algorithm works in noise-free, noisy, and very noisy signals for vowels as well as for voiced constants. The proposed pitch estimation only needs a very short frame length and gives accurate results at voice onset.<
>