Estimation of the vocal tract transfer function with application to glottal wave analysis

Abstract

A new method for determining the vocal tract transfer function for voiced speech is proposed. The method exploits the frequency domain characteristics of voiced speech and the concepts of minimum-phase systems. The short-time spectrum of voiced speech contains information about the glottal source and vocal tract filter components. In voiced speech there is usually a small frequency gap between the glottal source peak frequency response and the first formant of the vocal tract. The inherent inability of linear prediction parametric modelling in discriminating between closely spaced frequency components can make accurate modelling/estimation of the first formant (in the presence of a close and elevated peak due to the glottal flow) in voiced speech very difficult. A fixed pre-emphasis (single-pole, high-pass filter), commonly used in existing inverse filtering methods to reduce the influence of the glottal source, is not guaranteed to give the desired result across a range of voice types e.g. breathy, pressed and modal. The proposed method overcomes this problem by suppressing the glottal wave contribution using a dynamic, multi-pole, zero-phase lag high-pass filter, prior to analysis. In addition to minimising the influence of the glottal source, an expanded analysis region is provided in the form of a pseudo-closed phase. The technique also takes into cognizance the time varying nature of the vocal tract filter by determining, adaptively, an optimum vocal tract filter function using the properties of minimum phase systems. The performance of the new method is evaluated using synthesized and real speech. The results show that, under certain conditions, the method estimates the vocal tract formants and bandwidths more accurately than an existing closed phase inverse filtering technique.