Feature Analysis for Speaker Identification.

摘要

A method for text-independent speaker identification has been developed which utilizes vowel sounds as the basis for extracting speaker characteristics. The use of this approach typically requires that vowel samples first be classified according to vowel category, so that vowels of the same category can be compared in the speaker identification process. It has been demonstrated, however that it is only necessary to detect vowel-like sounds in the speech material and that speaker identification performance actually improves when there is no vowel recognition. Another significant outcome of this research was the successful application of sequential analysis to the decision process. The experiments which led to the conclusions of the research were based upon a data base of vowel samples for more than one hour of speech material excerpted from conversational speech recordings. Using the sampel 20-speaker reference patterns, tests with data from two speakers recorded two weeks after the reference material showed that one of the speakers was correctly identified in 18 seconds and the second speaker was correctly identified, but with a lower confidence level than the first.