Ten to twenty per cent of utterances collected by voice biometrics systems are not strong identifiers of the individual that spoke them, according to Dr. Clive Summerfield, the founder of Australian voice biometrics outfit Armorvox. Voice biometrics systems could therefore wrongly identify users under some circumstances.
Most voice biometrics implementations require users to utter a pass phrase or mention personal details as part of their authentication process. Dr. Summerfield told The Register that while a small fraction of the population, which he labels “wolves”, have voices that match many other voice prints, the need to know the pass phrase means voice biometrics systems are not likely to be casually cracked without an effort to also collect users' secret words. But he also feels that most voice biometrics systems build in tolerances for those with less distinct voice prints, therefore applying a lower authentication standard for all users.
Some of the less-effective voice prints are gathered because of ambient noise when utterances are collected. Signal clipping applied by carriers can also have the unintended consequence of reducing the quality of voice prints. Some individuals simply have generic voice prints that share qualities with many others. Summerfield labels those afflicted, for whatever reason, with poor voice prints as “goats”, in contrast with the majority of “sheep” whose voices are a strong authentication token.
Armorvox’s answer is a system it calls ImpostorMap which tests every utterance in a database to see if any could authenticate more than one user. Those with less-secure voiceprints can then be encouraged to re-enrol with a better sample. By doing so, Summerfield says voice biometrics can become a stronger authentication technique as users create more distinct utterance collections that are harder to imitate.
The company has already secured channel partners in Australia and is actively seeking implementation partners beyond Antipodean shores. ®