On February 6, at the RSA Conference in San Francisco, Andy Rolfe of Authentify and Dr. Stephen Elliot of Purdue University presented the results of a study on Authentify's implementation of voice biometrics/speaker verification. It should be noted that this study should not be considered to be a statement on voice biometrics on the whole, but on a specific implementation of a a single vendor's hosted authentication solution.

The Authentify system, as tested, offered a "high", "medium" and "low" security setting. The results follow:

Same Channel: Landline 

Same Channel: Mobile 

Cross Channel: Enroll on Landline, Verify on Mobile 

The Authentify solution uses the Nuance voice biometric engine (see information here ) and seems to only use numeric data for verification, which due to a small sample size is more problematic on cross channel, VoIP and mobile phone verifications.

Though the results are far from stellar, it does need to be stated that these results are from a single vendor implementation using the Nuance engine and is not representative of the entire industry. Results from different engines using larger amounts of speech data (words, sentences) will have significantly different results, as is shown by research published by the University of Canberra.

However, it is important to applaud Authentify for engaging with a trusted university to produce these results, as well as for publishing their information publically by presenting it at the RSA conference. We believe that the collection of more data can only help drive future of speaker verification. 

-- For point of disclosure, the author of this commentary does work in the voice biometrics field --