Psychometric Per Utterance Confusion Patterns In Listening Experiments

Andrew Lovitt, Electrical & Computer Engineering, UIUC

An analysis has been performed on the confusion matrices (CM) $P_{h|u}(\SNR)$ of a repeat of Miller-Nicely 1955 for 286 consonant-vowel (CV) utterances (16Cx1V), as a function of signal to noise ($\SNR$) (-18 to +30 dB). We show that for a given CV there is a large variability across utterances (i.e., talkers) as the $\SNR$ is decreased. Misarticulated utterances are dealt with by a simple entropy measure on the confusion matrix with no added noise. Many utterances are found to ``morph'' into a related sound (i.e., /ta/->/pa/) as the $\SNR$ is decreased. We believe that morphing is a useful tool in basic research on identify elemental perceptual speech features.