When the system invites customers to "tell me the reason for your call," the goal was to correctly interpret the response 65 percent of the time, and this figure has been exceeded, said Con Lambropoulos, general manager, customer access.
The objective is to improve customer service by routing callers to an appropriate agent as quickly as possible; "it was never about staff [numbers]," he said. Currently, Telstra answers 80 percent of calls within 60 seconds but aims to cut that to 20 seconds.
The project started by collecting the various utterances customers were using to describe what they wanted, including joke remarks such as "I want a pizza." A prototype was subjected to customer usability tests before the system went into limited production answering 10 percent of calls at the end of February.
Around half a dozen aspects needed some improvement before the trial was expanded in mid May to 25 percent of calls, but since May 29, all calls to 13 2200, 13 2000 and 125 111 have been handled by the system. A limited range of transactions including account balance enquiries can be completed within the speech recognition environment, but over 90 percent of calls are directed to an agent. That figure is in line with Telstra's expectation, as the primary goal was to allow customers to dial one of a small range of numbers and have their calls directed quickly to an appropriate person.
Customer feedback has been positive, said Lambropoulos, and customer service metrics have not been adversely affected by the introduction of the speech front end.
The system was developed in partnership with Dimension Data and uses the ScanSoft SpeakFreely software. Peter Chidiac, Asia Pacific regional director for ScanSoft's SpeechWorks division, said the software incorporates -millions of samples" of Australian speech with a variety of accents.