Microprocessor based access control technology!

Founded in 1999, Voice Security Systems Inc. has been involved in the development of voice/speaker verification/authorization systems using embedded technologies.

Two decades ago, securing access to devices, doors, services, etc. using speaker verification technology was still in its infancy. Processors were very slow when compared to today's modern electronics. Storage space was also at a premium. Many systems required prohibitively large storage facilities for verification software and voice template storage, even for a modest number of users. To resolve these problems, a method was developed which could adequately verify an individual's identity using their voice input, that did not require a large database of enrollment templates or sophisticated voice processing equipment.

The entire access control system which was developed by our founding engineers, included the storage of the user's voice data and the hardware and software required for the speaker verification process and fit into a standard double gang switch box (5"x5"x3").

Advances in electronics, DSP technology, processor speeds and storage facilities have allowed this once stand-alone, personal, access control system, to scale up to meet today's needs for secure access in an enterprise environment. The evolution of this technology is FreeKey®/Voice Protect®. Tomorrow's technology today !

This breakthrough technology also enables speaker verification features to be easily added to many existing microprocessor based devices such as PDAs, cell phones, computers, and smart cards. Most of these devices will not require the addition of any dedicated voice processing hardware or additional storage to utilize the FreeKey®/Voice Protect® method of speaker verification (a user's complete set of enrollment data fits in less than 800 bytes).

Some examples of applications for FreeKey®/Voice Protect® technology include:

Cell phone industry (to minimize losses from cell phone theft/fraud).

ATM manufacturers (to minimize losses from card and pin # fraud).

Automobile manufacturers (to dramatically reduce theft and car-jacking).

Some features unique to the FreeKey®/Voice Protect® method of speaker verification are:

Non-intrusive enrollment (takes less than 30 seconds).

FreeKey®/Voice Protect® has a very small storage requirement for enrollment data, which is easily stored on secure physical media such as smart cards, iButtons, or a consumer device such as a cell phone. The enrollment data is all that is needed, other than a live voice sample, to verify the claimed identity of the user.

The user can change their pass phrase more frequently to increase security (very much like text passwords are handled today).

The FreeKey®/Voice Protect® verification method adjusts for small physical changes in the user's voice over time, allowing long term use of enrollment data. In fact, the level of verification confidence can actually improve within the first several uses of the system.

System Performance

In response to several customer requests, we ported our technology to an Active-X prototype in order to provide an approximate comparison of our verification method against the verification products of some of the biggest players in the speech industry. In order to obtain comparable results, we choose the YOHO corpus speaker verification database as a basis for our testing in order to provide comparable accuracy measurements.

We are pleased to report the results of our testing showed very favorable results.

Out of one-hundred thirty-three users from the YOHO database we were able to enroll seventy-seven of them with an identical voice password constructed from the YOHO speech data.

Here is a summary of our test results using version 2.0.18 of the keypad test control available at this site:

Total Users Tested : 133
Total Users Enrolled : 77

Cohort Testing - Enrolled (77) vs. All (133 - one attempt each due to limited data)
Total False Accept (known phrase) : 94 / 10164 or < 0.93%

Verify Testing - Enrolled (66) vs. All (133 - 40 attempts each)
Total False Accept (unknown phrase) : 364 / 421652 or < 0.09%

Sixty-six of the users recordings were unable to generate satisfactory enrollment scores due to CO-articulation differences between the digits recorded. These user's enrollments were invalid, thus they were omitted from the Cohort Testing (where the user tries to impostor KNOWING the secret pass phrase). Data from ALL one-hundred thirty-three users was used for the verification testing to determine the "real-world" probability of someone trying to gain access to a stolen device for which the pass-phrase is unknown. The verify test phrases were all different from each other and different from the "correct" phrase that was compared to the impostor attempt. All of the phrases used consisted of "combination-lock" sequences; three two-digit numbers spoken as: "62...76...53". In an actual application, the user can create a "voice key" by saying virtually any phrase in any language as long as it meets a minimum energy requirement (about 2 seconds of speech). We believe this would further decrease the possibility of someone guessing the correct phrase, although the verification method still rejects most closely matched speaker(s) even if they know the phrase. Of course, these numbers will vary depending on the user, phrase selected and specific hardware used to extract the voice features..

For additional information about the test procedures, or to obtain information on how to run your own tests (from recorded data or live input) using the Keypad Active-X control, contact Voice Security Systems Inc. directly. We can also provide information on how to set up a persistent data store for the Active-X control. This allows "voice prints" to be saved on your hard disk for testing day-to-day performance and usability evaluation. We verify every day here at VSS. VSS "voice prints" that were created over a year ago are still functioning perfectly.

We believe we have a smaller, more robust and more efficient process of speaker authentication than anyone else currently in the industry. Our storage requirement for the "voice print" which is compared to a live sample is a little as 780 bytes (this is the actual requirement for the keypad demonstration on this site). The code needed to perform the validation is very compact and will fit into many inexpensive chips available today; including smart cards For example, adding the verification process to the Keypad Demonstration program only increased the program size by 2K. Those involved in adding speaker verification to the security layer of their application, are only now beginning to realize the benefits of a simple, unobtrusive user enrollment and verification process such as ours. Our method the first to operate without the need of a cohort set of enrollment data, requiring only three samples of a phrase chosen by the user. We were also the first to streamline enrollment procedure from several hours or even days, in some cases, down to only a few seconds. Creation of a new "voice key" (enrollment) can be completed in under 20 seconds. The process can run on as little as an 8-bit processor running at 8 MHz.