FEATURE EXTRACTION VIA THE BICEPSTRUM FOR RECOGNITION OF NOISY SPEECH
H FAIRHURST, C.C. Goodyear · 2024
Higher-order statistical techniques are attracting considerable current interest as a means for robust estimation of the parameters of noisy signals.For example, third-order cumulants are insensitive to added white or coloured Gaussian noise [1,2], and this insensitivity extends to other noise sources with symmetrical probability density functions These techniques therefore offer promise for improving the performance of speech recognisers when presented with noisy speech.A technique [2] that has been used previously is to obtain an overdetemiined set of linear equations from the third-order cumulant plane and apply aleast squares technique to derive the AR parameters at a chosen order.The power cepstral coe icients are then found from these and used to form the recognition feature vectors.The method requires choices to be made regarding the order of the AR model and the initial cumuiant equations, and the results may be sensitive to such choices.This paper describes the use of an alternative numerical technique that works for MA, AR or ARMA signals, does not require knowledge of the model or its order or the selection of starting equations.The method employs a two-dimensional deconvolutional technique [3] to compute the complex cepstmm from the third-order cumulants of the speech signal; from this an estimate of the power spectrum can be made.