NCJ Number
11428
Date Published
1972
Length
91 pages
Annotation
INVESTIGATIONS OF COMPUTER-AIDED IDENTIFICATIONS OF PERSONS USING 'SPOKEN WORDS' AS THE ONLY SOURCE OF IDENTIFICATION DATA.
Abstract
THE VOICE SPECTROGRAPH MACHINE HAS BEEN IN USE FOR SEVERAL YEARS BUT THE INHERENTLY SUBJECTIVE DECISIONS BASED ON ITS DATA ARE LESS THAN SATISFACTORY. ONE POSSIBLE SOLUTION TO THIS PROBLEM IS THE DEVELOPMENT OF A VOICE IDENTIFICATION MACHINE WHOSE PERFORMANCE WOULD BE COMPLETELY OBJECTIVE, REPEATABLE, AND VIRTUALLY INDEPENDENT OF THE SKILL OF THE OPERATOR. STANFORD RESEARCH INSTITUTE (SRI) MODIFIED A PREVIOUSLY DESIGNED COMPUTER-AIDED SPEECH ANALYSIS SYSTEM AND EVALUATED ITS APPLICABILITY TO SPEAKER IDENTIFICATION. THE RESULTS INDICATE THAT SPOKEN WORDS FROM THE SAME SPEAKER OR FROM DIFFERENT SPEAKERS CAN BE DETECTED WITH A HIGH DEGREE OF RELIABILITY WHERE THE RELIABILITY IS A FUNCTION OF THE DEGREE OF FEATURE MATCHING PRESCRIBED BY THE OPERATOR. THE PROTOTYPE SYSTEM IS FASTER AND REQUIRES MUCH LOWER OPERATOR SKILL LEVELS THAN METHODS USING VISUAL COMPARISONS OF SPECTROGRAMS. FOUR METHODS OF IMPROVING THE PRESENT PROTOTYPE SYSTEM ARE SUGGESTED FOR FURTHER INVESTIGATION AND IT IS RECOMMENDED THAT A PRACTICAL SEMIAUTOMATIC SPEAKER RECOGNITION SYSTEM BE DEVELOPED AND PLACED IN USE BY LAW ENFORCEMENT AGENCIES. AN ADDITIONAL STUDY RESULTING FROM A SUBCONTRACT LET TO TEXAS INSTRUMENTS PERFORMED SPEAKER IDENTIFICATION ANALYSES ON A LARGE DATA BASE PREVIOUSLY GENERATED BY COMPANY FUNDED SPEECH RESEARCH. BOTH CONTRACTORS PROVIDED STATISTICAL DATA ON EXPECTED IDENTIFICATION ERROR RATES OF THE ANALYSIS METHODS USED BY THEM. THIS PROJECT HAS SHOWN THAT A SEMI-AUTOMATIC SYSTEM IS A POSSIBLE APPROACH TO THE SPEAKER RECOGNITION PROBLEM. THE RESULTS OF THE SRI TESTING PROGRAM INDICATED THAT SUCH A SYSTEM WOULD HAVE AN ERROR RATE OF LESS THAN 1 PERCENT, PROVIDED NO DECISION WAS MADE 30 PERCENT OF THE TIME. (AUTHOR