NCJ Number
62294
Date Published
1979
Length
4 pages
Annotation
THE HUMAN SPEECH PRODUCTION APPARATUS IS EXAMINED ALONG WITH A SPEAKER VERIFICATION SYSTEMS AT PHILIPS RESEARCH LABORATORY IN THE FEDERAL REPUBLIC OF GERMANY.
Abstract
THE SYSTEM'S SIMULATED TWO-STAGE CLASSIFIER USES 391 BITS OF REFERENCE STORAGE ON A STANDARD MAGNETIC IDENTITY CARD. SPEAKER-CHARACTERISTIC FEATURES FED TO AN AUTOMATIC SPEAKER RECOGNITION SYSTEM MUST BE EXTRACTED FROM ACOUSTIC SPEECH SIGNALS. VARYING THE GEOMETRY OF THE VOCAL TRACT ALTERS ITS FREQUENCY CHARACTERISTICS AND PRODUCES DIFFERENT SPEECH SOUNDS. EXTRACTION OF THE VOCAL TRACT FREQUENCY CHARACTERISTICS AND DATA REDUCTION ARE IMPORTANT COMPONENTS TO THE AUTOMATIC SPEAKER VERIFICATION SYSTEM. IT IS TRAINED WITH SAMPLE UTTERANCES OF A CODE SENTENCE, AND THEN SPEAKER-SPECIFIC REFERENCE IS CALCULATED FROM SAMPLES AND STORED USING A MINICOMPUTER. TWO ASPECTS OF RELIABILITY MUST BE TAKEN INTO CONSIDERATION WITH RESPECT TO SPEAKER AUTHENTICATION SYSTEMS, SUCCESSFUL AUTHENTICATION TO ALL LEGITIMATE PERSONS IN ONE OR TWO ATTEMPTS, AND REFUSAL OF ALL NONLEGITIMATE PERSONS OR IMPOSTORS. WHEN TESTED, THE PHILIPS SYSTEM SHOWED A FALSE-REJECT RATE OF 0.67 PERCENT AND A FALSE-ACCEPT RATE OF 0.74 PERCENT FOR A SAMPLE OF 2,500 UTTERANCES BY 50 SPEAKERS. ILLUSTRATIONS AND REFERENCES ARE INCLUDED. (DEP)