Prediction of perceived phonetic distance from short-term spectra—a first step
Dennis H. Klatt · The Journal of the Acoustical Society of America · 1981
Some models of speech perception (and many speech recognition devices) posit a process whereby short-term speech spectra are compared with a set of reference templates. These schemes will succeed only to the extent that metrics can be found that are (1) sensitive to phonetically relevant spectral differences such as those caused by formant frequency changes, and (2) relatively insensitive to phonetically irrelevant spectral differences associated with a change in speaker identity or recording conditions. Judgements of phonetic distance between pairs of static synthetic vowels and fricatives have been collected in which the stimulus ensemble included formant frequency changes and a number of acoustic changes that turn out to have little phonetic relevance (e.g., spectral tilt, relative formant amplitudes, high-pass and low-pass filtering). These data can be used to evaluate a spectral distance metric. For example, distance calculations based on the sum-of-squares of differences in critical-band filter bank outputs and those based on the linear-prediction residual correlate poorly with the vowel distance-judgement data. A metric based on spectral slope differences near the peaks in the criticalband spectra to be compared can be made to correlate very well (0.95) with the perceptual data. Sensitivity of the correlation coefficient to several design parameters of this metric (such as number of critical-band filters, filter bandwidths, and aspects of the weighting function) will be discussed. [Work supported in part by an NSF grant.]