Speaker Identification by a Matching-from-Sample Technique

Abstract

This study examines the ability of observers to identify the speakers of short test utterance (monosyllabic words, spondee words, and phrases) when sample utterances by the speakers are available to the observers. In one series of experiments, utterances were presented aurally through headphones; in another series, the same utterances were presented visually as intensity-frequency-time patterns. The observer's task was to identify (or, in some experiments, to authenticate) the speaker of an utterance by making reference, as often as desired, to samples of that utterance by each of 8 known talkers, and to indicate the confidence of each judgment on a 4-point scale. Identification scores and confidence ratings were consistently higher for the aural than for the visual experiments. Error scores in the aural experiments seemed to be related to the phonetic content of the utterances. In the visual experiments, errors appeared to be dependent on both phonetic content and number of syllables. [Work supported by the U. S. Army Signal Corps; K. N. Stevens is also at the Massachusetts Institute of Technology, Cambridge, Massachusetts.]