Statistical Techniques for Talker Identification
01 April 1971
Many of us can perhaps recall the experience of identifying a caller on the telephone from a relatively short utterance such as the word "Hello." This might indicate that even short utterances contain sufficient information for identification, and it is an intriguing and interesting problem to inquire whether automatic, objective, accurate and economic methods can be developed for talker recognition. The authors of at least ten papers in the last eight years have reported experiments with (simulated) automatic talker recognizers. Using a variety of approaches to different aspects of the problem, these experimenters have met with strikingly similar success--90 percent (or more) correct recognition. Previous studies may be classified into two groups according to whether the problem addressed was verification (is the speaker who he claims to be?) or identification (assignment of an unknown utterance to one person in a given group of speakers). While two studies of the first kind involved 34 voices or less,1-2 the third 3 and most extensive 1427