AI can guess your age from the sound of your voice
By detecting the age of customer service callers, Nuance aims to prioritize senior citizens and prevent fraud.
By Jared Newman The next time you make a customer service call, your place in line could be determined by how old you are. Nuance, a company that makes voice recognition products for businesses, says it’s developed an age-detection AI that can single out senior citizens for priority service. It can also help detect fraud by looking for mismatches between the age of the caller and the account holder on file. Nuance’s first customer is the Spanish telecom giant Telefónica, which started asking for age-recognition tools to address a spike in customer service calls during the coronavirus pandemic. Within a month, Nuance had a solution in place, and now it’s hoping to sell the technology to banks, insurance companies, and government agencies as well. “We really showcased how with computer technology we can be far more precise and far more sophisticated than the human ear could be,” says Brett Beranek, the vice president and general manager of Nuance’s security and biometrics business. What the human ear misses Even without technological assistance, we humans do have some ability to determine people’s ages. Compare how a person sounds at 20 years old with how the same individual sounds at 80, for instance, and you may notice that the tone of voice develops a roughness with age. But without that frame of reference, pinpointing a speaker’s age gets trickier, especially for older speakers. A 2010 study from the Journal of Social, Evolutionary, and Cultural Psychology found that when people try to guess the age of male speakers over the age of 35, or female speakers over age 55, they’re at least 10 years off on average. They also tend to underestimate the age of speakers as they get older. Brett Beranek [Photo: courtesy of Nuance]Beranek says Nuance’s AI can be more accurate by detecting more than 1,000 “micro-characteristics” that the human ear can’t process or comprehend. “Unless somebody has an age-related disease that has a significant impact on the voice, it really isn’t a material change,” Beranek says. “It is somewhat subtle, much more subtle than between childhood and adulthood.” That’s not to say Nuance’s age detection is perfect. Beranek won’t disclose specific margins of error but says that as speakers approach senior age, the algorithm gets less accurate, with inaccuracy reaching the “double digits” for customers between the ages of 60 and 65. Beranek says that the Nuance is transparent about the technology’s limitations, and that its partners get to decide what level of inaccuracy is tolerable based on how they plan to use it. Singling out seniors Before the pandemic hit, Nuance was more interested in ignoring customers’ ages. Some of its biometric products, for instance, can authenticate users by the sound of their voice, so the algorithms must learn to tune out any subtle changes that occur as people get older. But when Telefónica came calling, Nuance realized it could flip the idea on its head. The carrier had been overwhelmed by customer service requests during Spain’s strict coronavirus lockdowns, and wanted to prioritize seniors. But its own attempts at determining callers’ ages weren’t working. Like most companies, Telefónica can use caller ID to look up customers’ account details (including their age), but that doesn’t work for new customers, or for ones whose phones are lost, damaged, or stolen. Seniors are very high-value targets for fraudsters. Fraudsters, on the other hand, tend not to be seniors.” Brett Beranek, Nuance “We did not want these people to be unconnected, because most of them are alone in their homes, and their only link with real life and their family is through their communications,” says Pedro Serrahima, Telefónica España’s private customers and multibrand business development director. Telefónica also tried asking customers to press a number on their phones if they were over age 65, but callers quickly figured out that they could game the system by lying. “Young people were saying they were old people, and were inventing stories,” Serrahima says. “They were disturbing the service for people who really needed it.” That led Telefónica to reach out about using age-detection algorithms as a solution. Nuance liked the idea so much that it fast-tracked a response in about four weeks. “It was a good cause,” Beranek says. “Here was an ability for us to potentially change the parameters and adapt our AI technology, and detect individuals 65 years and up.” As with other AI technologies, it’s possible for age detection to be abused. The same tools that allow a company to prioritize seniors could allow for discrimination. That’s also true for other forms of speech recognition that Nuance already offers, such as dialect and gender detection. Beranek says he’s aware of the risks but notes that Nuance’s technology isn’t available to just anyone. Customers have to work with Nuance’s speech scientists to set things up, and the company doesn’t work with repressive governments or support large-scale public surveillance. Nuance also pushes customers to adopt certain privacy practices, such as getting consent for voice recognition and providing ways to opt out, though these aren’t hard requirements. “We have made conscious decisions not to sell in certain geographic markets, and also in certain use cases that don’t align with our values,” he says. Beyond customer service Even as stay-at-home orders lift, Nuance believes age detection has broader utility. Telefónica wants to use the technology to fight fraud, and Nuance is marketing that same use case to banks and other financial firms. Older customers with money in retirement or savings accounts are frequent targets for identity theft, so age detection could help root out younger-sounding attackers. “Seniors are very high-value targets for fraudsters,” Beranek says. “Fraudsters, on the other hand, tend not to be seniors.” Beranek imagines a scenario where Nuance’s algorithms flag a caller that sounds younger than the account holder on file. A customer service representative could then ask for more information to verify that the call is legitimate. The company could also call the customer back to confirm any transactions that the original caller requested, without having to do so for every single request. “If we can help organizations fine-tune those suspicious calls, those that are highly likely to be perpetrated by fraudsters, a simple outbound call can determine if it really is fraudulent or not,” he says. Telefónica’s Pedro Serrahima already has other uses in mind. Detecting customers’ age, along with other factors such as mood or nationality, he says, could allow the company to pair callers with the ideal customer service representative for them. “Before this, it was a question of prioritization,” he says. “Now we hope to have better service for these kind of people.”