Phonetic information in the vowel spectrum: the meaning of mel-Frequency Cepstral Coefficients

Abstract

There is still disagreement in the acoustic phonetics literature on how phonetic information is encoded in the vowel acoustic spectrum. The “formant hypothesis” holds that formant frequency locations are the primary encoding of phonetic information. But perceptual experiments have shown that listeners can identify vowels, to a certain extent, even when formant peaks are suppressed. This has given rise to the “whole-spectrum” hypothesis, which describes each vowel segment in terms of a high-dimensional description of its entire spectrum. While the “whole-spectrum” hypothesis better predicts suppressed-formant vowel perception, one advantage of the “formant hypothesis” is that it parameterizes a vowel inventory of a language in terms of featural classes indexed by a few values of formant frequencies. These frequency scales serve to describe a language’s phonological organization and sound change. In this paper, we show that the mel-frequency Cepstral Coefficients (MFCCs), whole-spectrum parameterizations that have been used in speech technology from the 1970’s till today, also have a phonetic interpretation leading to the same featural classes as traditional description. This is despite the fact that for many decades they have been thought to not be interpretable. Our arguments are based on analyses of all vowel data from the TIMIT database, with large amounts of speaker, context, prosodic, and dialectal variability, using information theory, effect-size statistics, and Fourier theory. Our goal is to show that MFCCs can be useful for further developments in the field of acoustic phonetics, because while they extract phonetically-distinctive information from the entire spectrum, they can also further understanding of the linguistic structure of vowel spaces.