Acoustic Modeling and Feature Selection for Speech Recognition

Abstract

The investigation of the thesis can be divided into three parts. In the first part, a nonlinear dynamic system is proposed for formant tracking. Compared to previous formant trackers depending on least squares estimation of LPC coefficients, MUSIC (Multiple Signal Classification) and ESPRIT (Estimation of Signal Parameters via Rotational Invariance Techniques) are used to improve the accuracy of formant estimation. Furthermore, a mixture of nonlinear dynamic systems is developed to improve the performance of formant tracking. In the second part, the formant tracker system is extended to perform phoneme recognition. The results indicate that the incapability of estimating the system measurement error prevents the system from performing well in the phoneme recognition tasks. In the third part, an SVM and HMM combined system is used to prove that the formant information is indeed useful to distinguish different phonemes. And the result in this part suggests that the output of the SVM can be treated as a particular case of discriminant transformation of the original acoustic space and might be useful for speech recognition.