Linking natural speech recognition technology to transactional systems has long been thought a risky business, however Australia's largest multi-client service agency Centrelink says it finally has got the formula right - up to a point.

Over the last month, Centrelink claims to have migrated 30,000 clients, predominantly students required to report their casual earnings under the Working Credit initiative, onto a natural speech recognition platform which allows them to update their income status with the agency in almost real time.

"It automatically updates people's [client records] which are then processed [on a batch run] every 24 hours. We trialled it on students because they would be able to help identify what worked and what we needed to improve, and give us the sort of feedback we needed," said Centrelink national project support manager Anne Woods.

While the new system currently accounts for only 0.5 per cent of Centrelink's approximately six million clients, the agency claims the government saves two-fold. Firstly by making on-time and correct payments, which reduces error correction and reprocessing, and secondly by freeing up call centre staff to be deployed in areas where greater direct customer contact is needed.

The Centrelink voice solution is based on the ScanSoft OpenSpeech Recognizer platform, which operates as an additional software layer applied to the interactive voice response (IVR) functions of telephony network kits such as Genesys, Nortel or Cisco boxes. Rather than processing DTMF or touchtone signals, Recognizer pulls the response direct from voice, then pulls options and information from the back-end database of choice.

Even with such initial success, Centrelink management says it is acutely aware that speech recognition is not yet a universal solution. Woods said that services for clients who lacked English language skills - which for Centrelink is large, will be provided by people with skilled in those languages.

Although the Recognizer product has been successfully configured to different linguistic taxonomies overseas, Woods said multi-lingual speech recognition would only be explored if significant demand was demonstrated.

Centrelink general manager John Wadeson said that while voice recognition certainly cuts costs, the real gain was in improved service delivery and customer satisfaction, namely a huge drop off in complaints.

"We haven't really had any, which was a big surprise for us, but it's great. When we started [the Working Credit voice recognition trial], it had all the hallmarks of a major government agency failing its clients yet again. But the customers seem genuinely happy to use it which is just great," Wadeson beamed.