Robust formant tracking in noise

While many algorithms exist for accurate extraction of formant frequencies from a speech waveform, these algorithms are not typically shown to be robust in the presence of highly-transient background noise such as competing speech waveforms. Preliminary results are presented from an algorithm using time-varying adaptive filters that appears to be robust in the presence of white, Gaussian noise or a single competing speaker over a large range of signal-to-noise ratios (quiet to -6 dB). Use of a synthesized sentence, for which the actual formant frequencies are known, permits quantitative assessment of the algorithm's accuracy as a function of signal-to-noise ratio.