High pitch source isolation using complex cepstrum in the autocorrelation domain

This paper proposes a technique for improving the performance of linear predictive analysis by using homomorphic deconvolution. Estimating the vocal tract transfer function using linear prediction is often inaccurate when the speaker speaks with a high pitch. This is due to the convolution of the glottal output signal, which is a periodic impulse, with the vocal tract impulse response. We propose a method that will remove this convolution by using complex cepstrum homomorphic deconvolution in the autocorrelation domain. The technique has been tested for synthetic and natural speech of varying pitch frequency and has improved performance on formant frequency estimation than a similar technique proposed recently.