A new automated method for reliable speaker identification and verification over telephone channels

Notwithstanding the wide use of fully automatic method of feature extraction in speaker identification/verification tasks, based mainly on LPC, cepstrum and spectral band techniques, a recent work [1] by the authors demonstrated that in forensic applications semiautomatic, operator-assisted feature extraction procedures are more reliable. The drawbacks of a semiautomatic parameters extraction is either on the loss of real time or in the dipendence of the system performance on the skilled operator contribution. The proposed method of automatic parametrization of the speech signals seems to resolve these problems. Given the speech signal defined as a single realization of the speaker´s voice, an automatic vowel identifier based on energy, pitch and cepstrum information provides the first segmentation. Then a formant trajectories tracer is operated and a rule-based system selects the suitable and stable frames on the vowels avoiding double sounds and transients. The restricted population of vowel-like sound frames is then subjected to a clustering procedure on the F1/F2 plane when all the a-priori knowledge on the vowel sounds is a guide to select the best five clusters to assign to the five Italian vowels. After the vectorization of the whole talker set and of the unknown veices, if any, the Bayes classifier is invoked and the proper decision tests are performed. The results of this automated extraction/decision sequence are presented in the paper with comparison to the oparator dependent procedure.