Recognition Test for an Open Set of Talkers Represented by Their Acoustic Parameters

摘要

A method based on a precise statistical formulation is presented which allows the use of a limited set of acoustic parameters for the recognition of the speaker in an open test. A number of speech realizations are collected for a group of different speakers. A second collection is made of utterances to be compared with the known speakers. In the general case a new utterance may not belong to any of the talkers referred to the test. Error of false identification and rejection is estimated as a measure of procedure effectiveness. Some considerations are made about the feasibility of the proposed method in the field practice of speaker recognition.