An Improved Formant Tracking Algorithm Featuring Adaptive Pole Enhancement

The measurement of areas of spectral resonance (formants) ,in the.spectrum of speech is essential in the identification of voiced speech sounds in recognition systems employing knowledge-based feature extraction and interpretation.Measurement of formant centre frequency is crucial in the identification of vowels, and additionally, preand post-vocalic consonant identification is enhanced if formant movement, particularly that of the second formant (F2), can be tracked with respect to time.Linear prediction coding (LPC) techniques [1.2] offer a convenient method of spectral analysis of the speech waveform to allow formant estimation, out characteristically offer only moderate signal-to-noise immunity.with the limit of usefulness being at approximately lSdB signal-to noise power ratio (SNR).The novel technique presented in this paper employs spectral estimation using LPC techniques, but offers a substantial improvement in the signal-to-noise performance of formant frequency estimation.The objective of the technique is essentially to enhance the detection of singularities (poles) in the estimated transfer function of the vocal tract inherent in the LPC-based model.High noise iumunity in formant estimation is specifically conferred by the progressive over-estimation of model order in any single analysis frame, coupled with the use of off-axis spectral estimation.An averaging filter function is applied as the final step in each analysis frame to yield noise-robust estimates of formant centre frequency.