Speech Reconstituted from Spectra of Reduced Dimensionality: A Study of Intelligibility
K. P. Li, George W. Hughes, A. S. House · The Journal of the Acoustical Society of America · 1973
In this preliminary study, we attempt to relate the space defined by a statistical [principal component] analysis of speech spectra to perceptual space represented by intelligibility scores. Speech spectra were represented in terms of a varying number of dimensions [3, 5, 7, 10, 15, and 35], each corresponding to a different degree of statistical accuracy, and the waveform was reconstituted from the reduced spectral data by means of inverse FFT computation [K.-P. Li, G. W. Hughes, and A. S. House, J. Acoust. Soc. Amer. 49, 134(A) (1971)]. Consonantal intelligibility scores for isolated monosyllabic utterances by one talker were determined from the responses of 12 listeners to a total of 553 test items. The average score increases monotonically from 47% to 75% as the spectral dimensionality varies from 3 to 35. Along with the number of dimensions, manner and place of consonant articulation and the effects of the details of the reconstitution process enter into the relationship between statistical accuracy and intelligibility.