A formant extraction method using autocorrelation domain inverse filtering and focusing method

An analysis-by-synthesis formant extraction method which uses autocorrelation-domain inverse filtering (ADIFs) and multistep focusing is proposed. A cascade of second-order ADIFs, each of which corresponds to a formant, is used. This method significantly reduces computational cost, compared with the conventional analysis-by-synthesis formant extraction method which uses frequency-domain inverse filtering or time-domain inverse filtering. A human speech production model, specifically the glottal pole model, has been implemented in this system. Formant extraction experiments for synthesized speech and natural speech have proved the robustness of this method. The real-time formant system can be realized by a single chip signal processor