A new glottal flow pulses extraction approach based on jointly parametric and nonparametric estimation

Linear prediction (LP) and complex cepstrum (CC) approaches have been shown to be effective for extracting glottal flow pulses (GFPs). This paper proposes a new approach, which first employs an odd-order LP analysis to find parameters of an all-pole model by least-squares (LS) methods and then uses deconvolution to obtain coarse GFPs by inverse filtering. Then, CC-based phase decomposition is used to obtain a refined estimate by further suppressing the remaining minimum-phase information from results after the inverse filtering. We demonstrate the effectiveness of this new approach by applying it to both synthetic and real speech voiced utterances.