A new approach to formant estimation and modification based on pole interaction

In an LP model of speech, a speech formant is characterized by a pole frequency and its bandwidth. The conventional root-finding formant estimation algorithm derives the formant poles from the LP polynomial (coefficients) by selecting appropriate roots to simulate the formant resonances. However, the resulting formant polynomial suffers from the problem of spectral deviation since one or more "spurious" roots are deleted from the original LP polynomial. The main problem is pole interaction. We propose a new algorithm that characterizes speech formants by formant frequencies, as well as formant spectral densities, while formant bandwidths are modified to reduce the degradation of pole interaction. This approach is also applied to derive a formant modification algorithm that makes it possible to change the formant frequency while maintaining the desired formant structure. The experimental results show that the formant polynomial obtained by the proposed algorithm produces a more reliable formant spectrum than one that does not consider the pole interaction effect.