Showing papers on "Dynamic time warping published in 1967"

Automatic Speaker Authentication Using Speech Recognition Techniques

Abstract: While some degree of speaker recognition can be achieved using gross measures such as average glottal frequency or average spectrum, it seems certain that, for highly accurate speaker recognition or authentication, it will be necessary to know what is spoken. In applications where there is no control over what the speaker utters, automatic speech recognition becomes a prerequisite for automatic speaker authentication. Considerable success in automatic speech recognition has been achieved using analog‐threshold logic and this equipment has been used to select automatically portions of continuous speech for speaker authentication. Characteristics of continuous speech deviate substantially from those predicted by classical phonetics and many of these deviations can be used for speaker authentication. A number of different measures have been studied and automatically extracted. For vowels, the basic measure is the slope of the spectrum. For consonants, spectral characteristics, duration, and sequence of occurrences are useful measures. The over‐all accuracy of authentication can be quite high when results for many individual measurements are combined. Results for single and multiple measurements will be discussed. [Work sponsored by the U. S. Army Electronics Command.]