For companies with a consumer-facing retail operation, a major source of fraud in the contact center is a criminal pretending to be a new customer. The criminal uses a false identity to extract benefits, such as getting a new device under contract, and then abandons the contract and is not to be found. The criminal benefits to this behavior and similar activities such as account takeovers are so tempting that criminals organize into groups to systematically and repeatedly defraud a company, altering tactics whenever the company erects a new defense.
Voice Biometrics technology offers a powerful, undetectable defense that companies can implement at their contact center perimeter. Using recordings in the contact center, companies can apply voice biometrics to compare voices of new customers with voices from known fraudsters as well as other new customers. If the voice of the new customer matches a known fraudster or a recent previous new customer with a different identity, then the company can mark the conversation as highly likely to be fraudulent.
Voice biometrics like fingerprint and iris scans are not 100% perfect. Some fraud will still slip through and some false positives will occur, but if a majority of the fraud is correctly detected in advance that would not be otherwise, the reduction of losses can be dramatic.
Fraud Detection solutions that leverage voice biometrics work by using the power of SpeakFreely™ passive voice biometrics. VBG SpeakFreely™ constructs voiceprints of individuals from conversational audio. The source can be any conversational audio whether recorded or live streaming. So imagine submitting a portion of a call center conversation to the VBG SpeakFreely™ engine to create a voiceprint. The process of constructing a voiceprint from audio is a process called enrollment. Once an individual is enrolled, imagine submitting new samples of audio from the same or different person and requesting SpeakFreely™ to compare the new audio sample against that voiceprint to see if a match occurs.For example, suppose that based on history of being hit by fraud, you know of a specific phone call that was fraudulent and thus you have recordings of the person committing the fraud, regardless of who that person claimed to be at the time. You enroll the audio of this recording into the voice biometric engine and store it in the fraud database as “Fraudster-0001.” Then as diagrammed in Figure 1 below, new customer calls come in each hour, or each four hours, or each day, however frequently as you wish, you apply SpeakFreely™ to compare the recordings of the new customers against Fraudster-0001’s voiceprint. SpeakFreely™ gives you a score that indicates the likelihood of a match. A high score indicates potential fraud, thus giving you the opportunity to act before fulfilling what may be a fraudulent order. As represented in the figure below, by leveraging our Identification API request, you can make a single API request to compare a new sample against the entire database of known fraudsters.
A critical point to understand is that the audio sent to the voice biometric engine must be from ONLY one speaker, the person you want to enroll or the person you want to compare. Many contact centers can only store call center recordings in a monophonic format where both the caller and the agent are in a single track. Because the agent audio is also in the recording, the recording cannot be used to enroll the caller. Software exists that attempts to take a monophonic recording that is available from most call center recorders and separate speakers into different tracks for the purposes of speech analytics. This software, in our experience, will not work reliably enough for the purposes of automated fraud detection using voice biometrics.
What you need from your contact center recorder is a stereo recording where each speaker is in a different track but perfectly synchronized, as depicted in Figure 2. We then strip the one track of the caller from the stereo recording to send to the voice biometric engine. If your contact center does not have stereo recording capability, you may be able to capture audio using audio “tapping” equipment or inserting a passive stereo recorder into the call path.
The cost of Voice Biometrics for fraud depends on the number of comparisons of new customers against the quantity of voiceprints in the fraudster database, also considering how quickly you want results. The shorter the turnaround time, the more equipment required for parallel processing of all the voices and thus the higher the cost. Therefore, pricing a system depends greatly on your situation and will inevitably require a custom quote.
As an example, imagine a scenario with the following conditions:
For this scenario, you need to compare the 500 new customers each day against the 100 known fraudsters, all within 24 hours. 500 times 100 = 50,000 comparisons over 24 hours falls within a range where we do not need to add dedicated servers. And Cloud means fast setup and no installation fees. The pricing at this volume of transactions and voiceprints is on the order of 10 cents for each comparison and $15 per month per fraudster voiceprint. The total cost is therefore $50 per day for the comparisons, totaling approximately $1,500 per month for transactions, plus $1500 per month for voiceprints, equaling $3,000 per month, or equivalently $36,000 per year. This is only indicative pricing. Volume discounts, Enterprise Unlimited licensing, and On-Premises licensing are all available to adapt pricing to your situation.
A further consideration, for example, may be that you not only want to compare an audio sample against an existing database of fraudsters, but you also want to detect new fraudsters. New fraudsters will appear as a repeat new customer. So let’s say that you not only maintain a fraud database, but also you maintain a rolling database of new customer voiceprints for all customers who are new within the last two weeks. Voice Biometrics can tell you if a new customer under a specific name is also a new customer using a different name within the last two weeks. Maintaining a rolling database of new customers increases costs but also adds a layer of fraud prevention on top of fraud detection.
Finally, another option that adds complexity but may be highly desirable is to use live audio during the call as the comparison against the fraud database rather than waiting to process the recorded call after the fact. With live audio, you then have the possibility of sending screen pops to the agent screen and transferring the call to a special agent who knows how to handle a caller is highly likely a fraudster. This option is entirely possible but requires tighter integration into the call center part of the contact center.
If you have a business susceptible to fraud in the contact center, gather the information below to prepare for a consultation on applying voice biometrics to detect and prevent fraud: