Speech intelligibility measured by adaptive maximum-likelihood procedure

Abstract

This paper describes an adaptive maximum-likelihood procedure, originally developed for psychoacoustic measurements, applied to measure speech intelligibility using the modified rhyme test (MRT) [J. Acoust. Soc. Am. 37 (1965) 158]. The use of this adaptive procedure for estimating speech-to-noise ratio at various pre-selected word scores as an input parameter is presented. Listening tests conducted on subjects demonstrated that the required speech-to-noise ratio could be estimated with sufficient accuracy in an adaptive run consisting of fewer than 25 test items. Such a run length is also sufficient to maintain the relative frequency distribution of consonants present in the original MRT with an average error not exceeding 3% points. Depending on the target word score, different values of bias and standard deviation of level estimates are displayed by the adaptive procedure. These variables were analyzed using numerical simulations as well as the results of experiments, in which the signal-to-noise ratio was determined for a number of target word scores. Maximum-likelihood adaptive procedure is considered as a more efficient procedure than the other adaptive methods (e.g. staircase procedures). Owing to this efficiency, the maximum-likelihood adaptive procedure can be a useful tool for quick assessment of the speech-to-noise ratio of communication systems at various percent-correct word scores.