SPEAKER RCOGNITION STUDIES ON CHORAL SPEECH (FROM CARNAHAN CONFERENCE ON CRIME COUNTERMEASURES PROCEEDINGS, MAY 16-18, 1979, BY JOHN S JACKSON - SEE NCJ-62284)

• NCJ Number: 62306
• Author(s): R DUBES; A K JAIN; O I TOSI
• Date Published: 1979
• Length: 7 pages
• Grant Number(s): ENG 76-11936A01
• Sponsoring Agency: National Science Foundation
• Corporate Author: Ores Publications
• Publication Format: Document
• Language: English
• Country: United States of America
• Note: *This document is currently unavailable from NCJRS.

Annotation

TRANSLATING TEMPORAL SPEECH SAMPLES INTO A CHORAL SPECTRUM IN THESE SPEAKER RECOGNITION STUDIES HAD THE EFFECT OF MAKING SPEAKER REPRESENTATION INDEPENDENT OF LANGUAGE, TEXT, AND RECORDING EPOCH.

Abstract

IN SPEAKER IDENTIFICATION, AN UNKNOWN SPEAKER IS LINKED TO THE MOST SIMILAR SPEECH PATTERN IN A GIVEN POPULATION. IN SPEAKER VERIFICATION, THE DEGREE OF SIMILARITY BETWEEN SPEECH PATTERNS OF A GIVEN SPEAKER AND A CLAIMED IDENTITY IS ESTABLISHED. PROBLEMS IN SPEAKER RECOGNITION ARE MADE DIFFICULT, ESPECIALLY IN FORENSIC-RELATED SITUATIONS, BY RECORDING AND CONTENT FACTORS, THE ACTUAL TEXT, THE TIME AT WHICH A RECORDING IS MADE, AND SMALL SAMPLE SIZES. AS A WAY OF MINIMIZING THESE PROBLEMS, SAMPLES OF CONTINUOUS SPEECH WERE REPRESENTED BY CHORAL SPEECH PATTERNS. USING BOTH 40-SAMPLE AND 30-SAMPLE DATA SETS THAT INVOLVED READINGS BY MALE ADULT SUBJECTS, ANALYSTS EXAMINED CHORAL SPEECH PATTERNS BY CREATING A PROXIMITY MATRIX INDICATING RANK ORDERS OF SIMILARITIES BETWEEN ALL PAIRS OF PATTERNS AND THEN APPLYING HIERARCHICAL CLUSTERING TO IDENTIFY SIGNIFICANT SUBSETS OF PATTERNS. MULTIDIMENSIONAL SCALING AND PROJECTIONS WERE EMPLOYED TO ESTABLISH SIMPLE METRIC DATA REPRESENTATIONS. FINDINGS DEMONSTRATED THAT THE RECORDING ENVIRONMENT WAS CRUCIAL IN SPEAKER RECOGNITION BASED ON CHORAL SPEECH. THE STRONG CLUSTERING ACCORDING TO MODE OF RECORDING SUGGESTED THAT THE RECORDING CHANNEL WAS A MORE IMPORTANT FACTOR IN THE DATA STRUCTURE THAN SPEAKER IDENTITY, EVEN THOUGH SPEAKERS CLUSTERED WITHIN EACH CHANNEL. IT WAS DETERMINED THAT CHORAL SPEECH IS SENSITIVE TO DIFFERENCES IN SPEAKERS BUT INSENSITIVE TO TEXTS. THE CHORAL SPEECH PATTERN METHODOLOGY IS WELL-SUITED TO FORENSIC SITUATIONS BECAUSE THE STRUCTURE OF SMALL DATA SETS IN WHICH EACH SAMPLE CONTAINS A LARGE NUMBER OF FEATURES CAN BE STUDIED. METHODS TO COMPENSATE FOR THE EFFECT OF RECORDING CHANNEL, HOWEVER, NEED TO BE DEVISED, AND THE METHODOLOGY SHOULD BE TESTED UNDER A VARIETY OF TEST CONDITIONS TO VALIDATE ITS APPLICABILITY. FIGURES, TABLES, AND REFERENCES ARE PROVIDED. (DEP)