Publication | Closed Access
Test of an Automatic Speaker Verification Method with Intensively Trained Professional Mimics
21
Citations
0
References
1972
Year
Speech SciencesEngineeringSpeech KinematicsBiometricsSeveral Professional MimicsVoice EvaluationPhonologySpeech RecognitionSpeaker IdentificationSpeaker DiarizationRobust Speech RecognitionIntensive TrainingVoice RecognitionAcoustic AnalysisHealth SciencesTraining MethodSpeech SynthesisComputer ScienceSpeech AcousticSpeech AcquisitionSpeech TechnologySpeech CommunicationVoiceSpeech AcousticsSpeech ProcessingSpeech PerceptionLinguisticsSpeaker Recognition
Several professional mimics selected by audition from a large group, were given intensive training on utterances of the eight “customers” in Doddington's 40speaker population [J. Acoust. Soc. Amer. 49, 139 (A) (1971)]. The training method uses an interactive DDP-516 computer to provide the mimic with immediate playback of his practice utterances as well as immediate A-B comparison between his own and customer utterances. Recordings of the best utterances from the best four mimics were processed by the computer verification system described previously [Doddington, op cit., and R. C. Lummis, J. Acoust. Soc. Amer. 50, 106(A) (1971)]. The system uses five features for verification: pitch, level, first, second, and third formant frequencies. If the acceptance-rejection criterion that yields equal-error performance in Doddington's speaker population is used for the mimics, 27% of the best utterances by the best mimics would be accepted. This acceptance rate compares to 1.2% for nonmimicking impostors. Extensions to the verification procedure are being studied that will make successful mimicry more difficult.