Fast and robust formant detection from LP data

This paper introduces a method for real-time selective root finding from linear prediction (LP) coefficients using a combination of spectral peak picking and complex contour integration (CI). The proposed method locates roots within predefined areas of the complex z-plane, for instance roots which correspond to formants while other roots are ignored. It includes an approach to limit the search area (SEA) as much as possible. For this purpose, peaks of the group delay function (GDF) serve as pointers. A frequency weighted wGDF will be introduced in which a simple modification enables a parametric emphasis of the GDF spikes to separate merged formants. Thus, a nearly zero defected separation of peaks is possible even when these are very closely spaced. The performance and efficiency of the proposed wGDF-CI method is demonstrated by comparative error-analysis evaluated on a subset of the DARPA TIMIT corpus.

► Selective roots/formants from speech linear prediction data are located in real-time.
► Merged formants are separated by means of a modified group delay function.
► Formant extraction results are as good as results of formant trackers which base on standard root solver.
► The presented method is simple to implement on both floating and fixed-point digital signal processors.
► Several setting options make this method simple to adapt to specific application and system-requirements.