Voice in banking isn’t exactly a new concept. After all, we’ve engaged with bankers using our voices since the first borrowers approached Renaissance Italy’s 15th-century “mounts of piety” to ask for loans. When, 500 years later, telephone banking became an option, bank customers embraced that channel too.
But while using our voice to connect with banks isn’t new, digital voice technology has been accelerated by advances in automatic speaker recognition, text-to-speech and natural language processing, which together have substantially improved the accuracy rate of the technology. In 2012, the rate at which these systems inaccurately detected words was about 30 percent, says Vijay Balasubramaniyan, CEO of call center security firm Pindrop.
Today, Amazon, Google and Apple have gotten that error rate down to less than 5 percent. “A 5 percent error rate is human parity,” he explains. “You’re able to understand 19 out of 20 words and often you don’t need the 20th word; with context, you can get the 20th word. That’s a huge shift.”
Advanced digital voice technology—powered by artificial intelligence and machine learning—is making it even easier for us to access account information, conduct transactions and keep our accounts secure.
It’s you, it’s really you
“It’s really rare to add a security element that also improves the customer experience,” notes Brad Paige, president and CEO of Kennebunk Savings Bank, a $1.1 billion mutual in Kennebunk, Maine. Two-factor authentication, security questions—these all add extra steps (and often frustration) for customers. “Before, we spent more time identifying the customer than we did actually problem-solving,” Paige explains.
Kennebunk Savings recently launched a security feature that relies on today’s increased accuracy of voice recognition. The bank’s customers can opt in to use their own unique “voiceprint” to authenticate their account access. “Now we can get that ID out of the way very quickly at the front end of the call and get to what the customer really needs,” says Paige.
Here are the basics of how it works. A customer calls the bank and is offered an opportunity to enroll. It only takes 30 seconds of recording to create a unique voice print, and when the customer calls back, it takes less than three seconds for the system Kennebunk Savings uses to authenticate the caller’s voice. As the customer explains why they’re calling, the software analyzes the voice and gives the bank employee a signal for positive confirmation, a red signal for suspected fraud and a third sign to indicate a need for further authentication. The solution is targeted toward the bank’s small share of frequent callers—those who constantly needed to repeat security questions—and the bank’s metrics show adoption is growing, Paige explains.
The security measure also has a side benefit in fraud reduction, says Brett Beranek, VP and general manager for security at Nuance Communications, the company with which Kennebunk Savings partnered to deploy the system. It relies on language analysis to flag when unexpected vocabulary or sentence structure indicates a fraudster is trying to manipulate the system.
Within six months, this technology typically addresses 80 to 90 percent of losses in the contact center, Beranek says. It also specifically addresses fraud by customers’ family members—those who have the answers to security questions—and serves to prevent elder financial exploitation, Paige adds.
Ask and ye shall receive
But digital voice applications go well beyond account security. It’s providing new tools for bank customers to conduct transactions and get information quickly, without the hassle of calling or visiting a bank.
U.S. Bank was the first to offer voice banking on all three major digital assistants: Alexa, Google Home and Apple’s Siri. The work to develop these skills is based on R&D in natural language processing—converting voice to text—that began over seven years ago, says the company’s chief innovation officer, Dominic Venturo. That work was aimed at refining voice biometrics, but the lab pivoted to more ubiquitous fingerprint authentication (like Apple’s TouchID) and now facial authentication.
But that deep dive into NLP proved applicable to a variety of customer engagement technologies, from chatbots to digital personal assistants. Customers like using the technology to get “bite-size answers and information,” says Venturo. Current features include account balances; Alexa and Google Home also offer billing due dates and transaction details. And U.S. Bank customers can ask Alexa to pay their U.S. Bank credit card bill from a U.S. Bank deposit account.
While it’s not mainstream yet, the progress is undeniable. “There’s a decent list of things that we’re hearing folks want to be able to do,” Venturo adds. How do they know? In part, it’s because of requests of the digital assistant that the assistant can’t fulfill, which makes these technologies a key channel for real-time customer feedback. On the list of things customers would love: the capability to authorize payments to third parties. While there are security and risk factors to manage, “that’s an example of where we see a logical next step,” Venturo says.
At press time, 15 banks and credit unions—of various sizes from U.S. Bank down to $209 million First National Bank in Bellevue, Ohio—have Alexa skills available for download on Amazon. They allow customers to check loan and account balances, determine bill due dates, analyze spending patterns, find locations and more. First Citizens Bank in Raleigh, N.C., offers a mortgage-focused Alexa skill that allows users to ask questions about purchase planning, budgeting closing and mortgage terminology. Optum Bank, a Salt Lake City-based provider of health savings accounts and flex spending accounts, has a skill that answers HSA and FSA-focused queries. And then other banks offer their own AI-based tools interaction tools, such as Bank of America’s “Erica” chatbot that’s based in its mobile app. Ally Financial offers Ally Assist.
Avidia Bank, a $1.5 billion bank in Hudson, Mass., is working with its core processor FIS to develop and test Alexa skills. The test has not been without its challenges. One of the biggest: the pronunciation of the name, which Alexa was misunderstanding as “a video bank,” says Katelin Cwieka, AVP and social media and brand manager at the bank.
Two of Avidia’s branches have Echo devices that allow customers can test and demo the service. “I heard from our tech ambassador team that they have had a lot of success showing this to customers and it’s been a big wow factor, given that we are a community bank,” Cwieka adds.
“The reason these technologies have become so good is because of massive amounts of data,” says Balasubramaniyan. And that data poses its own risks. Smart devices have raised a variety of questions how well they secure and protect user data. Alexa has been the subject of news stories indicating that Amazon may be collecting and retaining more conversational data than users expect.
In May, a Washington Post writer reviewed the data collected over four years by his family’s Alexa device. He found “thousands of fragments of my life: spaghetti-timer requests, joking houseguests and random snippets of Downton Abbey. There were even sensitive conversations that somehow triggered Alexa’s ‘wake word’ to start recording, including my family discussing medication and a friend conducting a business deal.” Amazon contractors in Costa Rica, India and Romania were reported by Bloomberg to have access to voice data as part of efforts to improve Alexa’s performance.
To protect customers’ data, Avidia Bank’s skill relies on secure authentication via entering their passcode or biometric on the bank’s mobile app—meaning that Amazon doesn’t record a customer’s account or other personal information. “We share feedback and information with Amazon but we don’t share account information or things like that because it’s linked to the app,” Cwieka explains. “That encryption through the app is where that security feature comes in.” Meanwhile, U.S. Bank ensures that personally identifiable information isn’t available to people who don’t need to know it, and data is handled according to the same protocol as other sensitive data controlled by the bank, Venturo says.
What about the risk that voice data might be captured, recorded and reused by a fraudster to access accounts? “If we have five minutes to an hour of you speaking, we can actually get a machine to sound completely like you,” says Balasubramaniyan, emphasizing the need to encrypt and manage audio data appropriately. “As long as you’re storing just the voice print that’s one thing, but if you’re storing raw audio, you now have the person, their identifying information and their audio. These are things these kinds of organizations need to consider.”
However, voice technology is getting constantly better at detecting the nuances that make each one of our voices unique, Venturo says. “When you record something and then you play it back, there are some characteristics about that recording that a computer can tell are risk indicators that it may not be live.”
Synthetic voices may sound convincing to the human ear but not to computers, explains Beranek. “Our voices are naturally variable things, and if there’s one thing computers are not, it’s randomly variable.” Adding to the fraudster’s challenge is the need to mimic both language and speech, an incredibly complex set of interlocking variables.
But digital fakes are getting better. “It’s an arms race to make sure you stay ahead of it,” says Venturo. The liveness detection can be complemented by other techniques like multifactor authentication, and banks in this area are constantly evolving their methods to prevent fraud.
Voice technology is still young, and experts expect it to continue popping up in new places. Venturo anticipates seeing bank-enabled voice tools pop up on game consoles and in connected cars. “I’ve seen more change in in-car infotainment systems in the last three years than in a decade [before] in terms of the ability to provision payments and on-demand wifi,” he says. “It might be limited in what you let [customers] do for risk and security reasons, but I think that’s got to be on all the banks’ product road maps in order to keep up.”
“We want to be where our customers want us to be,” explains Brad Paige. “Obviously the big driver here is the security piece, but a huge driver for us as well was simply improving the customer experience.”
This article originally appeared in the July/August 2019 issue of the ABA Banking Journal and is reproduced here with permission ,