Automatic Speaker Verification Using Phoneme Spectra
G. D. Hair, Thomas W. Rekieta · The Journal of the Acoustical Society of America · 1972
A technique, amenable to completely automatic real-time application, for verifying the identity of a cooperative speaker has been demonstrated and evaluated. Wide-band digital recordings of 10 isolated words were made on each of 9 weeks for 230 speakers. Segments of selected phonemes were automatically edited and power-density spectra computed over a 7.5-kHz bandwidth. A preliminary analysis of 40 randomly selected speakers has been performed. Phoneme spectra for five repetitions were averaged to produce speaker standard pattern vectors with the remaining four repetitions used for error-rate estimation. Pattern vectors of dimension 180 (smoothed spectra of six phonemes), 26 (features derived from the spectra), and 206 (combination of spectra and features) were used for technique evaluation. Combined error rates of 2, 5, and 1%, respectively, were obtained using a simple hypersphere decision rule. It is concluded that reliable real-time speaker verification can be achieved with completely automatic digital processing techniques and modest pattern storage requirements.