Speech Intelligibility: A Model and an Experimental Study
Marilyn Sue Berman, Martin C. Schultz, Wilson P. Tanner · The Journal of the Acoustical Society of America · 1970
Studies in speech intelligibility have been constrained by the inability to predict the result of any alteration in a communication system. The conventional measure of percent correct is inadequate in that it deals with only some portion of the data, and is variably influenced by subject response bias, and in other significant ways. This paper offers a model of the speech communication process arising from statistical decision theory and information theory. An experimental testing of the model was undertaken. Four trained subjects responded to frequency-filtered speech samples in noise from two message sets. The first contained 12 single words and the other comprised 36 three-word sequences. A priori predictions on maximum performance for both lists were calculated from a model based on signal detectability. Sequence results were predicted from scores obtained on the single words according to the same model. A posteriori predictions were attempted by use of percentage-correct scores and measures derived from information theory. The paper supports the model proposed. [Research supported by a grant from the Office of Naval Research presented to the Sensory Intelligence Laboratory at the University of Michigan.]