NCJ Number
62290
Date Published
1979
Length
6 pages
Annotation
THIS PAPER DESCRIBES AN AUTOMATIC SPEAKER VERIFICATION SYSTEM THAT IS SUITABLE FOR DIFFERENT ENVIRONMENTAL CONDITIONS, AS WELL AS FEATURE SETS USED IN THE SYSTEM FOR VOICE CLASSIFICATION.
Abstract
SPEAKER VERIFICATION INVOLVES THE AUTHENTICATION OF A SPEAKER'S CLAIMED IDENTITY. THE AUTOMATIC SPEAKER VERIFICATION SYSTEM HAS BEEN ADAPTED FOR DATA GATHERING IN ORDER TO EVALUATE ITS PERFORMANCE. ANALOG SIGNALS ARE NOT ANALYZED IMMEDIATELY; RATHER, THEY ARE RECORDED ON TAPE. SUBSCRIBERS TO THE SYSTEM MUST ENTER A BOOTH AND USE A CUSTOMARY TELEPHONE. AFTER SUBSCRIBERS DIAL THEIR PERSONAL CODE NUMBER, THE SYSTEM ACKNOWLEDGES THEM AND GIVES ACOUSTICAL INSTRUCTIONS ON HOW TO USE THE SYSTEM. SPEAKERS ARE REQUESTED TO REPEAT PROMPTED CODE SENTENCES SEVEN TIMES. CODE SENTENCES CONTAIN SPEECH SOUNDS THAT HAVE BEEN PROVEN TO BE PARTICULARLY SUITABLE FOR SPEAKER RECOGNITION. TO EVALUATE THE AUTOMATIC SPEAKER VERIFICATION SYSTEM, FIVE MALE SPEAKERS WHO HAD SIMILAR VOICES WERE SELECTED BY SUBJECTIVE LISTENING TESTS. REFERENCE FILES WERE ESTABLISHED USING 10 LEARNING SAMPLES PER SPEAKER FROM A GERMAN CODE SENTENCE. EACH SPEAKER PRODUCED 60 'TRUE SPEAKER' TEST SAMPLES OVER A 2-WEEK PERIOD AND 60 'IMPOSTER' SAMPLES BY INTENTIONALLY IMITATING THE VOICES OF OTHER SPEAKERS. IMPOSTERS COULD LISTEN TO TAPE RECORDINGS AND ORIGINAL SPEECH SAMPLES OF THE OTHER SPEAKERS. ADDITIONALLY, IMPOSTERS WERE NOTIFIED ABOUT THE ACCEPTANCE OR REJECTION OF THEIR ATTEMPT AT DECEPTION AND THEY WERE INFORMED ABOUT THE VALUE OF THE SIMILARITY MEASURE. IN THIS WAY, IMPOSTERS WERE ABLE TO ESTIMATE THE INFLUENCE OF PARTICULAR VARIATIONS IN THEIR PRONUNCIATION. FEATURE SETS USED FOR VOICE CLASSIFICATION WERE SELECTED SHORT-TERM SPECTRA, LONG-TIME SPECTRA, INTENSITY CONTOUR, AND STATIONARITY CONTOUR. RESULTS OF THE VOICE RECOGNITION EXPERIMENT DEMONSTRATED THAT, BY COMBINING FEATURES, HIGH PERFORMANCE IN SPEAKER AUTHENTICATION IS ASSURED. ILLUSTRATIONS AND REFERENCES ARE INCLUDED. (DEP)