The intelligibility threshold level (ITL): A new approach for evaluating performance of digital speech communications processors
Caldwell P. Smith · The Journal of the Acoustical Society of America · 1979
Intelligibility performance of voice processors has been typically specified in average intelligibility scores: scores presumably equaled or exceeded by half the underlying population of scores, assuming a normal distribution. However, extensive multispeaker testing of a wide variety of processors over the past decade has shown conclusively that: (1) populations of scores for processors typically deviate significantly from normal distributions; (2) mean scores of individual talkers in any multispeaker test typically show highly significant differences (a = 0.001); and (3) significant differences in mean scores for phonetic features are also typical. To further an implied objective of intelligibility testing: estimation of the intelligibility level equaled or exceeded by, say 90% of a population of talkers and listeners, a new approach to evaluation has been established, in which the intelligibility data is analyzed to determine intelligibility threshold levels (ITL's): levels equaled or exceeded by specified fractions of the population of scores. The method, based on nonparametric statistics of the Kolmogorov-Smirnov type, involves rank-ordering the population of scores and establishing cumulative distribution functions and their confidence limits, from which ITL's and associated confidence levels can be readily ascertained. Comparisons of processors typically show larger variations in ITL's than in mean intelligibility scores.