From a sample of young male Californians, ten speakers were selected whose voices were approximately normally distributed with respect to the "easy-to-remember" versus "hard-toremember" judgments of a group of raters. A separate group of listeners each heard one of the voices, and, after delays of I, 2, or 4 weeks, tried to identify the voice they had heard, using an open-set, independent-judgment task. Distributions of the results did not differ from the distributions expected under the

more »

... sis of independent judgments. For both "heard previously" and "not heard previously" responses, there was a trend toward increasing accuracy as a function of increasing listener certainty. Overall, heard previously responses were less accurate than not heard previously responses. For heard previously responses, there was a trend toward decreasing accuracy as a function of delay between hearing a voice and trying to identify it. Information-theoretic analysis showed loss of information as a function of delay and provided means to quantify the effects of patterns of voice eonfusability. Signal-detection analysis revealed the similarity of results from diverse experimental paradigms. A "prototype" model is advanced to explain the fact that certain voices are preferentially selected as having been heard previously. The model also unites several previously unconnected findings in the literature on voice recognition and makes testable predictions. PACS numbers: 43.71.Bp INTRODUCTION This study addresses the question of how well people remember unfamiliar voices after delays of 1, 2, and 4 weeks and examines the processes underlying memory for voices. These issues are of both practical and theoretical importance since their answers bear on matters including the specifications required for speech storage and transmission systems, the validity of legal testimony involving identification by voice, and the general problem of auditory pattern recognition. Most previous studies of long-term memory for unfamiliar voices have used closed-set, multiple-choice formats (e.g., MeGehee, 1937, 1944; Clifford eta!., 1981; Saslovc and Yarmey, 1980; Legge eta!., 1984). In these tasks, listeners hear a target voice that they later attempt to select from among a set of voices in which they are truthfully informed that it appears once and only once. For example, in the earliest of these studies (McGehee, 1937 , listeners attempted to select a single target voice from a set of five male voices after delays that ranged from 1 day-5 months. Recognition scores declined from 83% after 1 day to 80.8% after 1 week, 68.5% after 2 weeks, 57% after l month, and to 13% after 5 months. McGehee provided no tests of the statistical significance of the decline, and subsequent studies have generally failed to show statistically significant differences between delay conditions in listeners' ability to identify voices over the delays studied here (see Clifford, 1980, for a review). In the 1937 study, McGehee used listener groups of differing sizes at the various delays, but she does not specify which groups were used at which delays. Moreover, she states that in two cases there was a "break in discipline" with "a spontaneous acclamation by the majority of students when they heard the repeating voice" (p. 259), but she does not specify in which cases or at which delays this occurred. Thompson (1985) used male voices in a six-voice lineup task in which listeners rated each voice as to whether it was the voice they had heard 1 week previously. They could also respond that the voice heard previously was not in the lineup or that they were not sure whether it was in the lineup. However, the listeners were not given the option of saying the voice heard previously was in the lineup more than once. Thus, from the viewpoint of the listeners, the experiment was an open-set task, but not an independent-judgment task. Such a task can be considered an open-set, multiple-choice task with a decision threshold imposed by the listener. The results were 62.1% correct identifications, 22.1% incorrect identifications, and 15.8% "not in lineup" or "not sure if inlineup" responses. The present study used an open-set, independent-judgment recognition task in which listeners each tried to remember a single voice. In the recognition phase of the experiment, the listeners were told that the voice that they heard previously might appear once, more than once, or not at all. They were, therefore, to make each judgment independently of all others. This task is more veridical to most realistic 913 J. Acoust. Sec. Am. 85 (2), February 1989 0001-4966/89/020913-13500.80 ¸ 1989 Acoustical Society of America 913 mated a normal distribution on hard-to-remember versus easy-to-remember ratings were selected by reference to a cumulative seminormal plot. Thus the voices that were selected are approximately normally distributed with respect to how difficult listeners believe they would be to remember, but are otherwise unspecified with respect to voice characteristics except that they are closely controlled for regional dialect, age, and sex, as described above. On the basis of the preceding results, three voices were selected as target voices: speaker 2, whose voice was judged by the combined rankings of all judges as next to easiest to remember; speaker 9, whose voice was judged next to hardest to remember; and speaker 5, whose voice was judged intermediate in difficulty. D. The experimental design A total of 90 listeners, all native speakers of English, were divided randomly into three groups of 30. Each of the three target voices was played to one of the three groups of 914 J. Acoust. Soc. Am., Vol. 85, No. 2. February 1989 Papcun et al.: Memory for voices 914