Estimation of formant parameters by sum-of-poles modeling
David H. Friedman · 2005
A method is proposed for estimation of speech formant parameters, based on a short-time (4-5 msec window) spectrum and its derivative with respect to location along the time axis. The vocal-tract impulse response is modeled as a sum of complex decaying exponentials, and the transfer function as a sum of pole terms, equivalent to a pole-zero model with zeros accounted for through freedom of the pole amplitudes. Formant frequencies are directly read out as zero-crossing locations of a spectral quotient function, and in some cases reasonable estimates of the formant bandwidths can also be obtained.