Optimizing the reliability of speech recognition scores

Speech recognition assessment involves a dilemma because clinicians want a test that is short and reliable, but statistical principles dictate that a short test is unreliable. Curves representing the variability of test scores based on the binomial model reveal that approximately 450 scorable items are needed in order to optimize the reliability of a speech recognition test. A testing approach was developed to achieve this sample size while retaining the principal features of the most commonly accepted speech recognition tests (i.e., monosyllabic words presented in an open-set format, verbal responses, and right/wrong scoring). It involves the use of an interactive computer program to present CNC words in 50 three-word groups, which are scored phonemically, resulting in 450 scorable items. Normal performance is described as a function of both presentation level and signal-to-noise ratio. Comparisons of test and retest scores for 100 individuals with normal hearing and 100 persons with sensorineural losses revealed that the approach achieves the degree of reliability predicted by the binomial model for both groups. Phoneme scores accounted for 99% of the variance of word scores for most of the performance range encountered in clinical practice, making it possible for test outcomes based on phonemic scoring to be expressed in terms of equivalent word recognition scores.