Speaker recognition

Date

1974

Journal Title

Journal ISSN

Volume Title

Publisher

Abstract

The speech spectra of a known group of ten speakers were used to identify a speaker arbitrarily selected from the original ten. The experimental work was restricted to a closed test. Five words were selected for the experiment. Audio data was recorded on a magnetic tape recorder, sampled and Fourier spectrum analyzed. Standard reference patterns of each speaker and of each word were made by using the averaging techniques. Majority Formant Selection Criteria, and All Formant Selection Criteria. Minimum Square Error Method and Maximum Correlation Coefficient Method were employed to compare each test (speech) sample of the unknown speaker with the reference patterns of all known speakers for the same word. An analysis of the experimental data showed that the Minimum Square Error Method yielded more accurate results than Maximum Correlation Coefficient Method. The correct identification of speakers for all words was obtained 74 percent of the time on the average, whereas for individual word it was 76 percent on the average. Distribution of error for individual words was not uniform. The spectrum of the individual word spoken by each speaker was similar. The present techniques employed are reasonably good for a limited number of samples but may need to be modified for a large number of samples.

Description

Keywords

Citation