Speaker identification from a minimal set of training data
Larry L. Pfeifer · The Journal of the Acoustical Society of America · 1975
Two different data bases have been prepared for the purpose of performing speaker identification experiments. The first data base consists of a set of 10 male talkers, speaking a list of well-controlled sentences. The second is a set of 20 talkers, 10 male and 10 female, speaking the rainbow passage. The experiments are carried out under what could be termed a “minimal information” basis, i.e., the speech utterances are bandlimited to 3200 Hz and there is only one training sample. Identification decisions are based upon the analysis of various sound units such as the vowels and nasals, and the use of a minimum-distance classifier. Speaker identification results are presented and discussed in terms of both open and closed-choice tests. Using the combination of both data bases, results are also presented on a text-independent experiment in which a 100% identification score was obtained for seven male speakers. [This research was supported by the Air Force Systems Command, Rome Air Development Center, Griffiss Air Force Base, New York, 13441.]