Adaptive linear prediction filtering for speech analysis

Abstract

An adaptive linear prediction filter utilizing Widrow's least-mean-square noisy gradient algorithm is under study as part of TRW's ongoing voice processing effort. The computational simplicity of the algorithm, as well as its ability to quickly adapt to changes in the statistics of the input speech, makes the algorithm attractive for voice processing. Unlike other block form LPC algorithms which process the data in frames at fixed update rates (autocorrelation and covariance methods), the filter coefficients are updated at every data sample and a new set of LPC parameters generated at any desired time (variable frame rate). Previous work with the adaptive filter has focused on a cascaded realization for formant analysis [L. B. Jackson and J. Bertrand, Proc. IEEE International Conference on Acoust. Speech Signal Process., pp. 84–86, April, 1976] or input signals other than voiced speech [L. J. Griffiths, IEEE Trans. Acoust. Speech Signal Process. ASSP-23, 207–222 (1975)]. Preliminary results suggest that the original algorithm as implemented by Griffiths may be unsuitable due to the presence of pitch pulses in voiced speech. A modified version of the algorithm using a fixed step size per iteration appears to minimize this problem resulting in better spectral estimates. Numerical examples which illustrate the properties of the algorithm as compared to other LPC algorithms currently used in speech analysis are included.