Formant data analysis from averaged short-time spectra

By visual inspection, pitch synchronously computed short-time power spectra show a significant maximum-minimum structure and a smooth variation in time. So those spectra seem to be helpful in automatic formant extraction by peak-picking or similar methods in vocoder systems. It is shown, however, that the segmentation errors of the device used for pitch determination lead to spectral peaks, wich are not closely related to to the slightly damped poles of the vocal tract. Therefore, the short-time spectrum of a fundamental period is computed by averaging the power spectra of gliding, Hanning weightened frames of fixed length over the duration of the current pitch period. The result is a smooth short-time power spectrum with peaks showing a significant relationship to the weakly damped poles or formants of the vocal tract. The consideration of the time varying spectra of shifted frames also gives some insight into the articulation process. This method might also be used for pitch extraction of bandlimited signals whose spectra do not contain the fundamental frequency itself.