Detection of nasalized vowels based on cepstra derived from the product spectrum

Abstract

Spectral representation of a signal becomes complete when both the magnitude and phase spectra are specified. Conventionaly, features for detecting nasalized vowel are derived considering magnitude spectrum only, ignoring phase spectrum. In this paper, a detection method for nasalized vowels are developed, where a product spectrum is defined that combines the information of both the magnitude and phase spectra. Unlike conventional mel-frequency cepstral co-efficients (MFCCs), MFCCs are derived from the product spectrum of the band-limited vowel, namely MFPSCCs and are fed to a linear discriminant analysis (LDA) based classifier for the detection of nasalized vowels from the mixture of nasalized and oral vowels. Simulation Results on TIMIT database show that for detecting nasalized vowels the proposed cepstral features derived from product spectrum outperform the cepstral features derived from the power spectrum in both clean and different noisy conditions.