Identical Twin Test Analysis using VoiceCheck™

Biometric authentication provides identification and access control to any restricted environment via either its physical or network and application infrastructure. It works based on physical characteristics of the human body such as voice, fingerprint, retina, iris scan, face detection etc. Voice biometrics is considered the most cost effective and efficient biometric method which carries in addition to patterns, the mood, character and other emotional aspects of the speaker.

Voice carries, in addition to patterns, the mood, character and emotional aspects of the speaker.

THC Technologies Corporation’s VoiceVantage™ VoiceCheck™ is one of the leading voice verification biometric tool available in the biometrics market. This tool has been developed based on years of signal processing research and is currently deployed in telecom, finance, business and law enforcement verticals. It is very cost effective and works well with common recording and telephony devices.

Deployed in

Voice verification also known as “Speaker Authentication” consists of several steps. It begins with enrollment, where the candidate provides multiple voice samples by pronouncing a phrase either given by the voice biometric system or chosen by the administrator. The next step is identity verification which verifies the provided live input against the prerecorded voice print to confirm identity. Verification is done by comparing the live voice sample with the enrolled voice print and calculating the comparison score based on the level of matching. A successful match will be the one with a comparison score equal to or above the specific threshold set by the administrator.

Comparing the live sample with the voice print consists of several steps; End Pointing, Analysis, Comparing and Scoring.

VoiceCheck™ uses a language-independent voice verification system in which a customer’s voice is enrolled into the system by creating and recording algorithmic examples of the customer’s speech (usually words or a text phrase) for use as a reference. When verifying an identity, the customer is prompted to speak the same words or phrase which is then compared to earlier stored versions. The two signals are compared based on signal timing, inflection and variance. If the new word or phrase is sufficiently similar to the previously stored versions, the claimed identity is accepted. If not, the person is assumed to be an imposter and the identity claim is rejected.

Identical twins are the most ideal candidates to validate any biometric system. Identical Twin testing is a superior reference and testing methodology as they are the closest natural imposters of each other. Non-identical twin tests have been used in the biometric industry to test and rate the efficiency of the biometric systems, however Identical Twin Testing is rare.

The requirement was to create multiple voice prints of the identical twins participating in the test. The first task was to select the recording hardware and software platforms. Two different recording hardware products were used in voice sample collection to add an additional variant to the testing. The next task was to select the appropriate voice collecting environment in which to conduct the identical twins test.

The criteria for the enrollment process and project guidelines were provided by the vendor in detail. The identified candidates were thoroughly briefed on the enrollment process and efforts were made to ensure they were aware of the expectations of the voice collection process.

Candidates were asked to record their name, day/month and state/country of birth and favorite color as part of each enrollment process. Each candidate performed the enrollment twice, once as themselves and the other time as the imposter of their counterpart. The Voice Collection tool provided by the vendor is capable of evaluating the quality of the samples in real time and directs the speaker to repeat their phrase in the case of poor quality.

Web version of Fig. 1 (the same output, as text)

,mitri_Anthony_R00,mitri_Anthony_R01,mitri_Anthony_R02,mitri_Anthony_…
mitri_Anthony_E00_V00_R00,100,58,65,58,24,0,0,17,
mitri_Anthony_E00_V00_R01,58,100,59,59,30,17,15,29,
mitri_Anthony_E00_V00_R02,65,59,100,62,31,9,3,31,
mitri_Anthony_E00_V00_R03,58,59,62,100,35,28,15,30,
mitri_Anthony_E00_V01_R00,100,48,64,47,57,74,63,59,
mitri_Anthony_E00_V01_R01,48,100,56,80,54,50,58,59,
mitri_Anthony_E00_V01_R02,69,56,100,53,58,62,65,65,
mitri_Anthony_E00_V01_R03,47,80,53,100,52,49,56,55,
mitri_Anthony_E00_V02_R00,100,56,59,53,49,45,36,43,
mitri_Anthony_E00_V02_R01,56,100,76,65,55,57,62,53,
mitri_Anthony_E00_V02_R02,59,76,100,65,56,55,60,57,
mitri_Anthony_E00_V02_R03,53,65,65,100,47,66,54,51,
mitri_Anthony_E00_V03_R00,100,88,72,73,0,0,0,0,
mitri_Anthony_E00_V03_R01,85,100,76,75,0,0,1,0,
mitri_Anthony_E00_V03_R02,72,76,100,90,0,0,0,0,
mitri_Anthony_E00_V03_R03,73,75,90,100,0,0,0,0,
,mitri_Alexis_R00,mitri_Alexis_R01,mitri_Alexis_R02,mitri_Alexis_R03,m…
mitri_Alexis_E00_V00_R00,100,37,42,34,5,23,14,16,
mitri_Alexis_E00_V00_R01,37,100,51,51,0,13,0,3,
mitri_Alexis_E00_V00_R02,42,51,100,46,0,25,24,21,
mitri_Alexis_E00_V00_R03,34,51,46,100,0,1,0,5,
mitri_Alexis_E00_V01_R00,100,62,49,51,53,47,37,48,
mitri_Alexis_E00_V01_R01,62,100,51,52,57,44,49,48,
mitri_Alexis_E00_V01_R02,49,51,100,47,33,53,45,32,
mitri_Alexis_E00_V01_R03,51,52,47,100,38,38,40,36,
mitri_Alexis_E00_V02_R00,100,59,45,48,47,52,36,39,
mitri_Alexis_E00_V02_R01,59,100,61,55,36,40,33,33,
mitri_Alexis_E00_V02_R02,45,61,100,62,39,41,37,42,
mitri_Alexis_E00_V02_R03,48,55,62,100,39,52,52,50,
mitri_Alexis_E00_V03_R00,100,56,37,26,0,0,0,0,
mitri_Alexis_E00_V03_R01,58,100,50,37,0,0,0,0,
mitri_Alexis_E00_V03_R02,37,50,100,62,0,0,0,0,
mitri_Alexis_E00_V03_R03,26,37,62,100,0,0,0,0,

All the recorded comparison scores from the verification process were statistically analyzed and displayed on a spreadsheet. The scores for each phrase were compared between the candidate’s score and imposter’s score per each specific recording device.

The lowest score of the Candidate with their own rendition is colored in green and the maximum imposter score is colored in red. An average score has been calculated for each phrase to summarize the findings.

Phrase 1 was the first and last name of the candidate. Four renditions of phrase 1 has been recorded from the candidate and the imposter using both of the recording devices.

For the twins, the last name is the same and the imposter has to imitate only the first name intentionally. The imposter scores are neither low nor high comparing to the other phrases.

Web version of Fig. 4. (same numbers as the paper)

Phrase 2 was the birth day and month of the candidate. Four renditions of phrase 2 were recorded from the candidate and from the imposter using both of the recording devices.

For the twins, phrase 2 was factually the same and the imposter didn’t need to imitate any part of it intentionally. The imposter scores are almost the same as phrase 3, at the same time relatively higher than phrase 1 and phrase 4.

Web version of Fig.5. (same numbers as the paper)

Phrase 3 was the birth city and state of the candidate. Four renditions of phrase 3 were recorded from the candidate and the imposter using both of the recording devices.

For the twins, phrase 3 is factually the same and the imposter didn’t need to imitate any part of it intentionally. The imposter scores are almost the same as phrase 2, at the same time relatively higher than phrase 1 and phrase 4.

Web version of Fig.6. (same numbers as the paper)

Phrase 4 was the favorite color of the candidate. Four renditions of phrase 4 were recorded from the candidate and the imposter using both the recording devices.

For the twins, phrase 4 is factually different between each other and the imposter needed to imitate the entire part of the phrase 4 intentionally. The imposter scores are remarkably low when compared to the other phrases.

Web version of Fig.7. (same numbers as the paper)

On analyzing the comparison scores generated by the VoiceCheck™ tool, it is clearly evident that the possibility of successfully spoofing between the identical twins is not possible provided the phrases are exclusively specific to the candidate. The comparison score of the imposter went above 60% but below 75% only twice among all the 32 attempts.

In both of these cases, the phrases are factually ‘true’ for the imposter. On average the comparison score is fairly high when the imposter has no need to lie or to make any effort to impersonate his counterpart. This clearly proves that voice is not just mere words; voice carries the emotional characteristics of the person at the moment of recording. Voice tone and difference in psychological intent, when speaking the truth versus attempting to impersonate, clearly caused a significant difference in the voice samples.

The scores for the phrase ‘favorite color’ which is factually different for each of the twins and the imposter score for this phrase were the absolute minimum when they attempted to spoof each other. This proves that choosing the phrase and its degree of exclusiveness to the individual is a key factor in effective and efficient authentication.

Sensitivity of the voice prints generated by VoiceCheck™’s algorithmic technology to all those above mentioned factors and properties of voice is very commendable and the results clearly prove that it is highly resistant to any such imposter attacks.