Evaluating spectral distance measures with reference to human perception
Yana Kane-Esrig, Lynn A. Streeter, C. Karore, Susan J. Devlin, Marian J. Macchi · The Journal of the Acoustical Society of America · 1987
Two important criteria for spectral distance measures in automatic speech recognition are: (1) the variance of distances between different tokens of the same utterance should be small and variances of distances between different utterances large, and (2) the pattern of distances across utterances should correlate with perceived phonetic similarity of the utterances. These criteria were used to evaluate several spectral distance measures, including (1) Euclidean distance between two log formant ratios, (2) LP-residual (“Itakura”) distance, (3) Manhattan distance between linearly spaced points on LP spectra, and (4) Manhattan distance between points on “perceptual” spectra (transforming the frequency scale to barks and convolving with an asymmetric filter of critical bandwidth). Distances were computed among synthetic utterances and among one speaker's natural utterances of 11 Dutch vowels produced in isolation. The perceptual similarity data were those reported by Pols, van der Kamp, and Plomp [J. Acoust. Soc. Am. 46, 458–467 (1969)]. Not surprisingly, Euclidean distances between log formant ratios predicted perceived similarity best, but this measure is of limited practical utility, since formant tracking is problematic. The second best measure for both evaluation criteria was distance between “perceptual” spectra. Distance calculations using these “perceptual” spectra are computationally feasible, produce the desired spread in intervowel distance distribution, and mimic perceived similarity.