A Speech Feature Vector based on its Maximum Phase Component

This paper examines the performance of a vowel classification scheme using a new form of feature vector derived from a decomposition of the speech segment into Maximum Phase and Minimum Phase components. Justification for this approach in terms of its perceptual relevance is first made, followed by a signal processing scheme to obtain the components. The form for the feature vector is then discussed. Lastly, experimental work compares the performance of this new feature vector under a variety of distortion conditions with the contemporary popular choice of Mel-Frequency Cepstral Coefficients