Maximum likelihood estimates (MLEs): An alternative approach for assessing speech intelligibility data
Caldwell P. Smith · The Journal of the Acoustical Society of America · 1990
The Diagnostic Rhyme Test (ANSI S3.2-1989) for assessing speech intelligibility results in arrays of scores for phonetic features, talkers, and processor conditions that can readily be converted to arrays of frequency counts representing numbers of correct responses by listeners. This transformation creates an opportunity for analyzing intelligibility data as contingency tables and applying the methods of discrete multivariate analysis to separately assess effects contributed by phonetic features, by talkers, and by processor conditions. The method is one of testing hypotheses: Particular combinations of main effects and interactions are postulated, and maximum likelihood estimates (MLEs) for each cell in the array under those constraints are calculated. The model is compared with the source array by means of chi-squared or G-squared measures, with the result providing a basis for rejecting or not rejecting the hypothesized independence from those effects omitted in the model. A pilot study was conducted in which this method was applied in comparisons of speech intelligibility from various processing conditions, as well as data from replicated tests. Results were compared with those from traditional analysis of variance, and with the use of the t test for paired data items.