Noise robust formant frequency estimation method for speech recognition based on spectral model of repeated autocorrelation of speech

Abstract

Formants frequencies of the voiced utterance represent the free resonances of the human vocal tract system. They are one of the fundamental properties of human voiced speech, and for the purpose of speech analysis or speech recognition, formant frequencies play a dominant role. In this thesis, effective methods for formant estimation are developed, which work well even in the presence of significant background noise. In real life applica- tions, very often human speech is affected by environmental noises from different sources. Hence noise robustness of formant estimation methods is a key factor. Accurate estima- tion of formants from given noise corrupted speech is a very difficult task. The major objective of this research is to develop an algorithm that can successfully estimate the formants in the presence of noise, overcoming the limitations of conventional methods. The autocorrelation operation on the speech signal can be viewed as a mean to overcome the adverse effects of noise, since it offers advantageous property of strengthening the dominant formant peaks, leading to better formant estimation accuracy in noise. One major idea in this research, unlike the conventional spectral domain peak picking is to develop a spectral model of autocorrelated speech signal and thereby introduce a model fitting scheme to find out the model parameters which are directly related to formants. Based on the spectral peak strengthening property of the autocorrelation operation by introducing new poles on the formant location, the idea of repeated autocorrelation is presented. The effects of repeated autocorrelation in time and frequency domains are investigated in detail, especially in noisy environments. It is observed that that in com- parison to single autocorrelation, double autocorrelation function of a signal exhibits more noise immunity. A spectral model is further developed to incorporate the effects of double autocorrelation. Finally the effect of spectral band limiting of the speech signal before performing the autocorrelation operation is investigated. It is shown that formant estimation from each band further improves the estimation performance. In order to utilize this property, a band limiting approach is developed that can adaptively filter the frequency zones where a formant frequency is most likely to be present. Spectral model for the double autocorrelation function of the band limited signal is proposed and em- ployed in a model matching approach for estimating the formants. Several vowel sounds taken from the naturally spoken continuous speech signal are tested in the presence of noise. Vowel sounds from synthetic speech as well as naturally spoken isolated words are also considered. The experimental results demonstrate superior performance obtained by the proposed scheme in comparison to some of the existing methods at low levels of signal-to-noise ratio. The estimated formants are used in a basic vowel recognition scheme utilizing a linear discriminant analysis based classifier along with Mel frequency cepstral coefficients (MFCC), and the results demonstrate a good degree of noise robust- ness compared to the methods using formant values estimated using traditional formant estimation schemes.