Prediction of the Course of F2 across Consonants

Abstract

In a formant-tracking vocoder, any lag in acquiring correct formant-frequency values, after the release of a consonant, produces false formant transitions; these may cause an error in transmission of the consonant. Tracking of the second formant (F2) is especially critical in this respect. In the present study, some means of predicting F2 are presented. F2 relations across consonant constrictions were measured in spectrograms of the spontaneous speech of five male speakers. The study compared prediction of the frequency of postconsonantal F2 by three extrapolations of preconsonantal F2: (1) holding the last value; (2) linear extrapolation of the last course; and (3) exponential extrapolation of the last course. Holding and exponential extrapolation were superior to linear extrapolation. Mean prediction error was about 300 cps with the two better methods. Errors were then partitioned according to upward or downward last course, in conjunction with high-middle, or low-frequency regions for the beginning of extrapolation. Certain partitionings reduced mean prediction error to 200 cps. [Research supported in part by the U. S. Army Electronics Research and Development Laboratory.]