Automatic Formant Tracking

A formant tracker has been implemented which assigns frequency, amplitude, and bandwidth to each of the first three formants for each 10-msec voiced segment of continuous speech. Its input parameters are fundamental frequency, rms, and up to five spectral peaks below 5000 Hz with their respective amplitudes and bandwidths. Peak information is obtained from linear prediction spectra. Techniques distinguishing this formant tracker from previously reported ones are as follows: (1) All spectrum computations are accomplished in the peak picking phase, before ferment tracking—this may yield spurious formants but few missing ferments. (2) Anchor points are located by selecting three consecutive 10-msec segments in which three formant frequencies do not differ from one segment to the next by more than a threshold amount. Tracks for the first three ferments are extended in both directions from these anchor points. (3) Frequency pattern matching aids decision making when more than one ferment is possible for a given slot; this is particularly useful for nasals and /l/ and /w/. Frequency pattern information is derived from speaker vowel-sonorant frequency tables. [This research was supported by the Advanced Research Projects Agency of the Department of Defense under Contract No. DAHC15-73-C-0080.]