Negative binomial probability models for speech intelligibility data
Caldwell P. Smith · The Journal of the Acoustical Society of America · 1986
Since reporting at ASA*50 on success in fitting negative binomial probability distributions to speech intelligibility data, extensive additional analyses have been made on Diagnostic Rhyme Test intelligibility data from multi-speaker tests. Modeling involves converting feature scores to frequency counts of listener errors, for each speaker's scores. Scores for the four primary feature states for each of the six phonetic attributes of the DRT represent outcomes of presenting eight tokens to eight listeners; thus error frequency counts from zero to 64 are possible, for each of the 24 attribute states. After conversion of scores to frequencies, a cumulative distribution was formed from the ranked data, a “frequency of frequencies” distribution. The mean and variance of these distributions provided parameters for negative binomial distributions, which were compared with chi-squared and Kolmogorov-Smirnov tests. Similar tests were performed on sub-sets of the intelligibility data representing partitioning into scores for individual speakers, for all male speakers, all female speakers, voiced feature scores, and unvoiced feature scores. Analysis of intelligibility data from tests of a variety of conditions revealed that in the majority of cases these distributions did not differ significantly from the negative binomial probability models, either for the composite data or the partitioned data sets. These findings open up new avenues for assessment and interpretations of speech intelligibility data.