Formant estimation with linear prediction
Walter C. Gish · The Journal of the Acoustical Society of America · 1980
Although the vocal tract transfer function for vowels is approximately all-pole, the resulting vowel waveform is not. We have investigated the effects of using an all-pole analysis technique (linear prediction) on real and synthetic vowels. Formant estimation accuracy and pole location stability were studied as a function of the length and placement of the speech segment analyzed, its pitch, and the analysis technique used. The synthetic speech consisted of a glottal waveform model [G. Fant, STL-QPSR 1 (1979)] with time varying formant damping. Covariance and autocorrelation linear prediction was performed over closed glottis phases, entire pitch periods, and randomly placed 20–25 ms frames. Closed glottis phases for real speech were identified from simultaneous laryngograph measurements and verified by inverse filtering. Covariance linear prediction over the closed glottis phase gave the most accurate formant estimates, stable pole locations, and little or no degradation with increasing pitch. Other analysis intervals or techniques gave significantly poorer results. [Work supported by NSF.]