Measurement of consonant confusions by human listeners in whispered speech

Abstract

The acoustic differences between whispered and normally phonated speech are large; yet whispered speech is still very discriminable. Since replacing a periodic voicing source with an aperiodic excitation does not significantly affect perception, there exists the possibility of a central processing stage to augment the normal peripheral processing. In the present experiments, the 16 CV syllables used by Miller and Nicely (1955) are whispered by three male and three female talkers. The data are used to collect confusion matrices from listeners who discriminate among 64 sample syllables from each talker. Data are also collected with stimuli involving normal and loud speech, in an attempt to determine what, if any, acoustic cues might discriminate between the consonants across stimulus conditions. The consonant confusions are obtained using a multitalker babble as a masker. Babble is used instead of white noise because it should equally mask all speech stimuli, whereas white noise may mask some stimuli, such as fricatives, more than others. Probabilities of consonant confusions across the voicing dimension for whispered speech will be presented.