Experiments with speaker verification over the telephone

SUMMARY Wehavepresented aseries ofexperiments inspeaker ver-ification for both high quality speech and telephone speechusing a statistical approach based on HMM phone mod-els. The decoding procedure has been efficiently imple-mentedbyprocessingallthemodelsinparallelusingatime-synchronous beam search strategy. Speaker verification (oridentification) can be carried out in both text-dependent ortext-independent modes using the same phone models.For text-independent verification, the phone based ap-proach was shown to clearly out-perform a simpler Gaus-sian mixture model on high-quality speech from the BREFcorpus and fortelephone speech has a 20% lower a posteri-oriequal errorrate. FortheBREF corpus, text-independentand text-dependentverificationEERs wereaboutthesame. 2 On the telephone corpus, text-dependent verification per-forms better than text-independent. When a verificationattempt fails, allowing a second trial reduces the number oferrors by 20%, while only increasing the number of trialsby 10%. For the telephone speech corpus, the majority ofthe errors are due to low scores for a few target speakers,mostly reflecting differences in the origin of the call for th etraining and testing sessions. In an additional experimenton the telephone speech, combining a model for F0 withthe speaker-specific phone model set did not significantlyimprove performance. On the telephone speech corpus, ana posteriori equal error rate of 2.9% was obtained using aminimum duration of 2s per trial, in text-dependent mode,allowing 2 trials per attempt. This can be contrasted withthe equal error rate obtained on the high quality speechcorpus which is well under 1%.