Systematic errors in the formant analysis of steady-state vowels. (English) Zbl 1005.68805

Summary: The locations of formants in a speech signal are usually estimated by computing the linear predictive coefficients (LPC) over a sliding window and finding the peaks in the spectrum of the resulting LP filter. The peak locations are estimated either by root-solving or by computing a coarse spectrum and finding its maxima. We discuss four sources of systematic error in this analysis: (1) quantization of the speech signal due to the fundamental frequency, (2) incorrect order for the LP filter, (3) exclusive reliance upon root-solving, and (4) the three-point parabolic interpolation used to compensate for the coarse spectrum. We show that the expected error due to F0 quantization is  10% of F0, and that the other three sources can independently skew the final formant estimates by 10-80 Hz. We also show that errors due to incorrect filter order are related to systematic differences between speakers and phonetic classes, and that root-solving is especially error-prone for low formants or when formants are close to each other. We discuss methods for avoiding these errors and improving the accuracy of formant estimation, and give a heuristic for estimating the optimal filter order of a steady-state signal.


68U99 Computing methodologies and applications
68T10 Pattern recognition, speech recognition
Full Text: DOI