Influence of pitch on formant frequencies and bandwidths obtained by linear prediction analysis

As is well known, the predictor coefficients are obtained by minimizing the mean‐squared prediction error over a speech segment typically 10 to 20 msec long. For periodic signals, such as voiced speech, the prediction error is large at least once during a pitch period. It is shown that the inclusion of speech samples with large prediction error in the analysis interval introduces errors in the predictor coefficients and in the formant frequencies and bandwidths derived from them. However, these errors are small if the vocal‐tract resonances die out effectively within a pitch period. This is often not the case, particularly for the first formant. The errors can be significantly reduced by either confining the analysis to the closed‐glottis part of the pitch period or by eliminating those samples from the analysis which have large prediction error. Methods for determining predictor parameters accurately from the autocorrelation function of a periodic signal are also presented.