Acoustic invariance in speech: Simple or complex properties?

Abstract

Recent research has suggested that invariant acoustic properties characterized by the shape of the short-time spectrum at consonantal release can be derived for place of articulation in stop consonants. Templates were developed to capture these properties and they successfully categorized 83% of stop consonants spoken by different speakers and varying in vowel contexts [S. E. Blumstein and K. N. Stevens, J. Acoust. Soc. Am. 62, 526 (1977)]. However, in addition to their characteristic shape, each of the three properties seems to have a characteristic frequency prominence with low, mid, and high energy peaks corresponding to labial, velar, and alveolar consonants, respectively. Thus, acoustic invariance might be derived from a simple property sensitive to the frequency in which the highest energy peak of the spectrum falls, rather than from a more complex property sensitive to the overall shape of the spectrum. To investigate this question, the same spectra for initial stop consonants analyzed in terms of templates characterizing their overall spectral shape were reanalyzed in terms of the location of their prominent energy peak. Unlike the templates characterizing shape, the frequency template categorized correctly only 53% of the utterances; vowel context effects were found for particular consonants; and overlap between the alveolar and velar categories was found in the frequency range 3000–3500 Hz. Implications for a theory of speech perception are discussed. [Work supported by NIH.]