A Speech Channel Evaluation Divorced from Talker-Listener Influence

This paper shows how a speech channel can be evaluated so that the characteristics of the talker-listener group, as well as associated electrical and acoustical subchannels, do not influence the final result. The characteristics of the talker and the listener are assumed to remain invariant with respect to time. Although perfect invariance is a mathematical idealization, the assumption of invariance is a good one if the talker is kept unaware of the listener's response and the channel's behavior, and if listener learning is minimized, for example, by the use of phonemes or nonsense syllables. Speech fidelity is defined by the transition matrix, which is the array of conditional probabilities of the output symbols when the input symbol is specified. The method presented demands the confusion matrix display of the results of a two-part articulation test performed by the talkerlistener group. The first part establishes the transition matrix of the generalized talker and of the generalized listener and is performed with the speech channel replaced by a perfect channel in which no errors, or confusions, can occur; the second part establishes the transition matrix of the speech channel and is performed upon the speech channel. The transition matrix of the speech channel is obtained by premultiplying the confusion matrix resulting from the second part of the test by the inverse of the talker's transition matrix, and then postmultiplying by the inverse of the listener's transition matrix. The data supporting the method have been sufficient to show that the transition matrix of the speech channel is insensitive to the division of error between the talker and the listener. Thus, the listener may be charged with all the error so that the transition matrix of the speech channel is given by the second confusion matrix postmultiplied by the inverse of the first confusion matrix.