Application of Distribution-Free Tolerance Regions to Automatic Speaker Verification
Guy W. Beakley, F. B. Tuteur · The Journal of the Acoustical Society of America · 1969
The effectiveness of different methods for constructing distribution-free tolerance regions for classification purposes is investigated in a speaker-verification problem. It is assumed that much of the information for the recognition of speakers is contained in the transition between phonemes as well as in the phonemes themselves. For this reason, many short-term spectra of readings of the word my were calculated. These spectra were used to form three different 256-dimensional measurement spaces. The spaces differed in the number of spectra used and the coarseness of each spectrum. Various methods were employed for constructing distribution-free tolerance regions with fixed expected false-alarm probabilities. The recognition capabilities of these methods were tested in the different measurement spaces. Other decision procedures were also tested, and the results were compared. Error rates for several proposed hypersphere tolerance region procedures ranged from 1.99% to 5.3% for 40 main-speaker training samples, 208 impostor training samples, 185 main-speaker test samples, and 439 impostor test samples.