Learning Center
Frequently Asked Questions
- What is voice biometrics?
- What is a voice print?
- What is the difference between speaker verification and speech recognition?
- Who develops the voice engine you use?
- How does enrollment work?
- How does verification work?
- What is a "normal voice"?
- What if I have a cold?
- What about twins?
- Can I use any type of phone?
- What about thresholds?
- Can voice settings be customized for individual users?
- What are FAR and FRR?
- Is voice verification 100% secure?
- Does a voice print need much storage space?
- Who is currently using voice verification as a means to secure confidential data?
- Where can I get more information?
1. What is voice biometrics?
Voice biometrics refers to the process of comparing a digitized sample of a person's voice with a stored model 'voice print'. Like other biometrics such as fingerprints, it has been established that an individual's voice is unique and can therefore be safely used to prove one's identity.
Back to top
2. What is a voice print?
A voice print is a digital representation of the unique features of an individual's voice. Voices differ due to a number of physiological characteristics such as the vocal chords, trachea, nasal passages, how the tongue moves inside one's mouth to produce certain sounds, etc. The combination of these characteristics is analyzed and identified as unique for everyone. A voice print is not a sound file (recording) of the individual and therefore cannot be replayed as a recording by an impostor.
Back to top
3. What is the difference between speaker verification and speech recognition?
Speech recognition is concerned with WHAT is being said; the system listens to the words that are spoken by the user and acts upon the spoken command by giving the user the option requested. Speaker verification which is concerned with WHO said it. The system listens to how the user speaks, and decides whether the user is who he or she claims to be. The voice engine used by MEK is language and vocabulary independent. The user can say anything in the language of choice, which makes it very user-friendly and ideal for international deployment.
Back to top
4. Who develops the voice engine you use?
MEK uses market-leading voice biometrics technology including the ScanSoft, Inc. Speech Secure™ voice engine, which was acquired with the purchase of the SpeechWorks organization in 2002. This voice verification engine was originally owned by T-NETIX, who developed it based on patented technology from Rutgers University. Speech scientists who developed the original engine have relocated to continue their work with it, resulting in ScanSoft having the industry's most experienced developers on their team.
Back to top
5. How does enrollment work?
The initial enrollment process only takes a few minutes. The system prompts the user to answer up to three specific yet simple questions, for example your first, middle and last name or the month and year you were born. The answers form the passphrases, which will be used when later accessing secure data. Remember, it doesn't matter what you say, but how you say it; any questions can be prompted, so long as the answer is one a user will never forget. For each question, the user provides three samples of his or her answer to the questions. The samples must be, at minimum, three syllables and one second of speech in order to create a robust voice print. The three samples are superimposed, background noise is extracted and the voice print is created within a few seconds. The system then proceeds to the next question and so forth (additional security requires multiple voice prints). After a few minutes, voice prints have been created for the user, which are used each time the secured system is accessed.
Back to top
6. How does verification work?
When users access secure data, their voice is compared to the previously stored voice print models. In MEK's help desk applications, the user is randomly prompted for at least two of the three passphrases. If both passphrases provided pass the biometrics test, the user is authenticated and is given a menu of automated options. If one of the passphrases is not accepted, the system prompts for the third passphrase which, if accepted, allows the user further access. If the system is not confident of the authenticity of the user after three passphrases, the user is denied access and can be sent to a live operator, or the connection can simply be dropped.
Back to top
7. What is a "normal voice"?
As with other biometric applications, success in using a voice verification system lies in providing constant sampling. In the field of fingerprinting, this means not having cuts or dirty fingers, for example. With voice biometrics, consistent sampling means speaking normally, in a relaxed manner, such as when you are introducing yourself. Users must recognize that conditions like being out of breath, chewing gum and drinking will affect the voice when verifying.
Back to top
8. What if I have a cold?
Not all of the parts that make up your voice template are affected by a cold; the voice verification engine can still determine who you are with a common cold. Serious throat illnesses, such as laryngitis, do require help desk assistance.
Back to top
9. What about twins?
To date, evaluation of twins has never produced an exact voice print match, despite valiant impersonation attempts!
Back to top
10. Can I use any type of phone?
Different phones produce energy differences (channels) which greatly affect how the voice is captured and used by the authentication engine. Using different channels will produce significantly different voice comparison results. As a safeguard, MEK provided users with multiple voice prints, in order for them to enroll on up to three different types of phone. The system then transparently and automatically attempts to find a match by comparing the voice sample to each set of voice prints enrolled by the user.
Back to top
11. What about thresholds?
The process of voice verification (comparing an utterance against a stored voice print) produces a score which indicates how close of a match the spoken passphrase compares to the stored voice print model. For simplicity, the MEK verification process produces a score between -10,000 and +10,000. In theory, a score of 0 or below indicates "likely an impostor"; while a score above 0 indicates "likely a valid user". In order to ensure the highest level of security whilst keeping the system user-friendly, MEK provides the ability to establish a "minimal threshold" at both the system and individual user level. The specified threshold signifies the minimum score that must be met in order to be recognized as the true identity. It should be recognized that threshold settings affect both the false acceptance rate (FAR) and false reject rate (FRR) in an inversely proportional manner. For example, a high threshold is a strong impostor deterrent but requires a more precise valid user.
Back to top
12. Can voice settings be customized for individual users?
In order to ensure the highest level of security whilst keeping the system user-friendly, thresholds can be altered for individuals who require a specific lower or higher threshold. The administrator can assign both a system-wide threshold as well as individual thresholds, for optimal performance. MEK can also customize the application, restricting access based on the time of day, number of accesses, group restrictions and other security features your organization may require.
Back to top
13. What are FAR and FRR?
The measure of performance by which the system will be accepted is tailored for each organization and application environment. Often, organizations will declare that no impostors should ever go through (False Acceptance Rate) and that no more than x% of valid calls should be rejected (False Reject Rate). Realistic numbers should be observed in recognizing that no system is 100% accurate and that FAR and FRR are affected correspondingly when the other is modified. For example, reducing the FAR will generally raise the FRR proportionately. Prior training of the users will considerably affect the performance of the system. MEK's experience shows that one should expect, in a standard production environment, that in order to achieve an FAR of "0% or as close as possible", an FRR of 5-8% should be expected. In a password reset application, these rates translate to the example that if a user forgot his/her password every month, over the space of a year, (s)he might be asked to repeat the third passphrase because the system did not recognize the user after hearing the first two, as is usually the case.
Back to top
14. Is voice verification 100% secure?
No solution, including biometrics, can de declared as 100% secure. However, MEK has implemented a number of features that make it virtually impossible to successfully assume another user's identity.
Back to top
15. Does a voice print need much storage space?
Depending on the length of the sample, you should currently allow for an average of 25KB per voice print.
Back to top
16. Who is currently using voice verification as a means to secure confidential data?
MEK currently has applications deployed all over the world. Wireless and telecom companies are using voice verification to secure long distance lines. Fortune 500 companies have chosen voice to secure automated password reset. Government agencies are using voice to aid in PKI deployment and administration, and to protect their vital and sensitive information and resources. Financial institutions are using MEK's voice biometrics applications for a multitude of unique custom projects.
Back to top
17. Where can I get more information?
Contact us to find out how we can solve your business challenges.
Back to top
|
