The encoding of vowel features in Mel-Frequency Cepstral Coefficients

Abstract

Most work on acoustic phonetics uses formant frequencies as the parameterization of the phonetic signal for understanding the acoustic difference between the sounds of the world’s languages. Work in speech technology, however, has relied for several decades on Linear Prediction Coefficients (LPC) and Mel-Frequency Cepstral Coefficients (MFCC’s), due to their greater invariance to physical differences between speakers. This paper explores the phonetics of the MFCC’s, asking whether these coefficients can be used by phoneticians to develop a greater understanding of the phonetic nature of speech segments. This is done through an analysis of the ability of individual coefficients to distinguish between American English vowels in the Hillenbrand database.