Speech recognition performs best in high-volume environments where customers are conducting loads of standard transactions that require call centre operators to enter a single, common set of details.
For New Zealand's Auckland Co-op Taxis (ACT), the technology has proved to be a lifesaver by dramatically slashing the number of calls handled by humans.
With 700 fleet vehicles and a regular customer base of more than 2000 corporations, ACT has always been a keen adopter of new technologies. One of its most useful recent investments was the use of Caller Line Identification information to identify callers and automatically pull up details such as their name and pickup address.
Such innovation is critical to keep on top of the company's burgeoning business: its 75 call centre staff handle around 3.7 million calls per year across seven New Zealand cities, and its taxis carry seven million passengers yearly.
In the highly competitive taxi industry, success is largely determined by the number of calls that can get through. Aiming to reduce its cost per call and increase the number of calls it could handle, earlier this year ACT began investigating speech recognition as an alternative to manual call handling.
"Cost per call is pretty high from a labour perspective, and it doesn't take long to add up," says Robert van Heiningen, ACT's chairman and CEO. "The industry suffers severely from peaks and troughs, and it's very difficult to ensure you've got sufficient staff sitting around to handle peaks without wasting their time during troughs."
In May, ACT went live with VeCab, a voice-enabled taxi booking service from speech recognition provider VeCommerce. The system, which is built on top of VeCommerce's FirstContact platform, was installed on a pair of Compaq DL380 servers, which run VeCab and the Nuance voice recognition engine. Aculab Prosody 2 telephony cards provide an interface with ACT's Lucent PABX, and VeCommerce also developed a special interface between the speech server and ACT's Raywood booking system.
An automated taxi booking service requires the ability to read back address and suburb details. This presents a big problem using prerecorded scripts, since it would theoretically be necessary to record someone speaking the name of every street in New Zealand. To get around this problem, ACT and VeCommerce complemented the system with a text-to-speech engine that allows the computer to phonetically pronounce any text it's fed.
New Zealand's dual-language culture added to the complexity of that application. Many Maori names are pronounced much differently than they're spelt in English, which means the text-to-speech software would make a horrible mess of them. To get around this problem, ACT implemented an innovative text-to-speech engine that provides proper pronunciation of both English and Maori words.
In its earliest form, VeCab was only used to let callers confirm that it had identified the right pickup details using the CLI system. Instead of pushing the number "1" to confirm the information, customers were asked to say the words "I'm ready now". Even with this limited application, takeup doubled from 17 percent of callers to 33 percent.
After about three months spent getting customers used to the concept of speech recognition, ACT expanded the system to include natural language recognition. This now means customers can say the time they want the car to pick them up, at any point up to seven days in the future. Customers can also speak to the system to request a special vehicle such as a van or a pair of cars.
Throughout the process, ACT and VeCommerce have recorded and analysed many customer sessions to see how they interact with the system. The results have been used to tweak the system's interface and improve the overall experience for customers.
"All the technology you can imagine is out there already," says van Heiningen. "It's got unlimited potential. The trouble is, we're trying to educate a couple of million people. You have to do that slowly, or customers get frustrated and they won't ring you again." Thanks to its caution during the rollout, ACT's system has been a resounding success. It currently handles some 53 percent of all bookings--more than twice the company's initial expectations--and will pay for itself within its first year of operation.
Customers, who could previously wait three minutes or more during peak times, are now transferred through immediately, reducing the number of people who hang up in frustration. Overall, the company is now handling twice as many calls as it was five years ago.
Despite initial staff concerns that speech recognition would make them redundant, van Heiningen says staff have warmed to its possibilities and relish the ability to focus on more interesting work instead of the monotony of basic bookings. No staff have been shed as a result of the new technology; in fact, employee churn has dropped significantly.
"People are sticking with us longer than before, which is great because a lot of training goes into them," says van Heiningen. "The feedback has been astronomical, because people don't like hanging on the phone. The gap between us and our competitors is getting bigger all the time. There are no losers in this one."
