Interactive voice response (IVR) has been helping computers interact with humans since a machine known as Voder became the first device to demonstrate electronically synthesized human speech in 1939.
IVR systems have changed dramatically in the intervening years. The DTMF (Dual-Tone Multi-Frequency) keypad menu has dominated the landscape since touch-tone dialing became commercially available in the 1960s, and was the system of choice for telecommunications company TalkTalk until just a few years ago.
TalkTalk had expanded exponentially through a series of acquisitions and organic growth, leaving an increasingly complex customer base. Consumer behaviour was being transformed simultaneously, with customers increasingly expecting smooth communications in whatever channel they chose.
Navigating the multiple layers of options in the DTMP service was no longer adequate. The company wanted to replace it with a contact channel that was simple, reliable and effortless for customers and added levels of automation to expedite their experience. TalkTalk turned to Nuance Communications to find the solution.
The language technology company provides intelligent self-service solutions supporting more than 14 billion customer engagements per year, as part of its modest ambition to "reinvent the relationship between humans and machines using AI". It provides platforms to companies such as Ford, Barclays and HMRC, and a number of telecommunications companies including AT&T, Verizon, and T-Mobile.
The collaboration with TalkTalk resulted in the implementation of the Nuance Natural Language Call Steering within the telecoms company's IVR.
The system offers intelligent conversation to customers and reduces the automated layers they need to navigate. It combines statistical models, intelligent dialogues and best practices of design to predict conversational intent in order to answer queries and complete transactions at rapid speeds.
Adoption rates have been high, but if clients don’t want to interact with the IVR, they can be guided through to an agent or fall back to a conventional DTMF menu.
"The interaction feels to a customer that it's free-flowing but guided to a solution," says Sarah Bramwell, who heads up the IVR and Telephony team at TalkTalk. "It’s not in any shape or form to reduce or stop our customers speaking to an agent. It's just to assist."
A recent addition called Smart Adaptation automatically analyses each voice interaction so that the product is continuously learning and improving its accuracy.
"All of our services are now natural language," says Bramwell. "We have a recognition rate of 95% which is one of the highest recognition rates of natural language solutions. We're continually optimising that solution."
The company claims it has reduced its customers' time spent on the phone by more than 18,000 hours over two years using the system, saving an average of 26 seconds per call. Call costs have dropped by more than 10 percent each year, and time spent within the IVR by 25 percent due to customers being consistently and immediately routed to the correct agent.
The system has now reached a speech recognition rate of 95 percent, and an ROI was achieved within a year of the rollout.
TalkTalk has also used Nuance's voice biometrics as a replacement for traditional passwords. In the year since it was rolled out, 900,000 customers have signed up.
"They just have to say: 'With TalkSafe my voice is my password'," says Bramwell. "You're prompted to say it so you don't need to remember anything, and then you're straight into the conversation with the person who's going to help you with your query. We've saved 60 seconds in terms of time that a customer spends with an agent as a result of them verifying from their voice."
"What TalkTalk is using is called text-dependent voice biometrics, where you're in an automated system," adds Nuance's director of voice biometrics product strategy Brett Beranek.
Organisations have the option to customise the specific phrase used to gain access, but Nuance doesn't recommend using personal passwords for individual customers, as it adds another password to forget but with little security. What matters is not what is said, but the voice that is saying it.
The company also produces a system of text-independent voice biometrics that operates during a live conversation with an agent, analysing the customer's voice as they're speaking. This can be a better fit for situations where a human conversation is essential, such as contact centres dealing with sensitive information.
"Voice biometrics is somewhat different than the other biometric modalities that you're probably familiar with," says Beranek. "Fingerprint, iris, DNA testing, all this is just 100 percent physical characteristics.
"Voice biometrics is measuring some of those physical characteristics. Not all of those are static. The shape and size of our larynx changes over time, the voice passes through our nasal passages, bounces of our teeth, so your teeth grow, you lose a tooth, all that will change the characteristics of your voice.
"So some of those physical characteristics are dynamic, but we're also measuring up behavioral characteristics."
They include accents, the rhythm of speech, how syllables are spaced out, and whether the customers speak in a monotone pitch or in peaks and valleys. Together they form a unique voiceprint, the audio equivalent of an individual fingerprint. If a customer has a cold the system can rule that the print has still been sufficiently emulated to proceed as normal.
If a whole series of characteristics are off then a further form of verification may be required to validate that person's identity and avoid fraud. Any ambitions of mimicry are muffled by algorithms that can detect whether a voice is recorded due to small signatures that are left in the audio. The voiceprint is left as evidence at the scene of the crime.
Stealing a database of voiceprints is of little worth to a fraudster, as they’re extremely difficult to replicate artificially, but it does have value if they can delete the voiceprints. Security measures such as firewalls and encryption help to reduce this risk.
The company has also added behavioural biometrics to track customer interactions with their device, such as how they tap their phone or type with a keyboard, to prevent fraudsters using malware to bypass the security system once the authentication is completed.
Bramwell doesn't have any extra demands of the technology, but will use the benefits it provides to further develop the customer service at TalkTalk.
"What I’m looking at is because now we have 900,000 customers who are enrolled for voice biometrics and TalkSafe, that allows me to start to scope out what automation and self-serve strategies I can build that would actually today require ID and verification," she says.
Services such as adding a TV channel that would normally require human password verification in order to access the customer account could be conducted simply and securely through a self-serve model if customers already have an enrolled voiceprint.
Nuance does have some new innovations on their way, such as a drive to combine voice with other biometric modalities that will be backed up by a risk engine. Customers want choice, and incorporating the more discreet facial recognition could provide a method of accessing personal information such as their account balance.
The downside of choice is that customers may not always opt for the safest option. Nuance has developed a solution: a risk engine that allows customers to only use the less secure modalities for lower-risk interactions.
The IVR and biometrics have been a success across TalkTalk’s broad demographic of customers. The company’s experience is supported by a study commissioned by Nuance that suggested there was no measurable difference in adoption of voice biometrics between a 75-year-old and a 20-year-old.
"I always assumed that biometrics in general would be more adopted by younger generations, but I was proven wrong,” says Beranek.
"If the individual sees a huge improvement in the convenience, and they perceive the technology as being secure as well, then the adoption will occur."