Coping with peaks of the share market can play havoc with call centre staffing.
On quiet days, call agents can sit around waiting for the phone to ring, while busy days see queues of investors lining up on the switchboard to place orders.
Which is why Commonwealth Securities (ComSec) has been quietly trialling a natural language voice recognition system it hopes will shoulder some of the load.
"When the market is busy, it is very difficult to handle peaks in call flow," Paul Rickard, managing director of ComSec, said.
"While our Internet site takes some of the load off our customer representatives, they are still snowed under at times. At the same time, we can't afford to have 50 people waiting for the phone to ring."The system, called VoiceBroker, will handle those calls, as well as taking after-hours trades. About 10 per cent of the stockbroker's 520,000 clients are testing VoiceBroker now. The system provides stock prices only now, but the trading firm expects to have transaction function by its official launch in September.
Rickard told Computerworld one of the main advantages in implementing the voice recognition technology was the fact ComSec had no existing IVR (interactive voice response) system in place, which is traditionally used by companies to direct callers to a specific business function such as paying bills, and the like. This system works well, unless there is a huge sample of alternatives for customers to wade through, such as all the stock codes, Rickard explained.
"It was to our advantage to start from scratch as there were no integration issues involved," he added.
The system was developed in Australia by Syrinx Speech Systems, and was chosen for its ability to provide continuous, speaker-independent speech recognition.
For example, in ComSec's case, it was imperative that the system recognises the difference between similar sounding words, such as 'Keycorp' and 'ecorp', to avoid costly mistakes in placing trades for clients.
Syrinx' speech recognition technology is speaker independent, in that it is phonetically-based said Clive Summerfield, Syrinx' CEO. It is also able to interpret accents, background noises and speech patterns.
"Firstly we developed a dialogue for ComSec, which is a conversation interface between a person and computer," Summerfield said.
"Then we created a vocabulary unique to ComSec, which would include stock names and codes."The company also optimised the system's recogniser for most of the target population's speech patterns, which is the Australian accent.
The technology responds to continuous speech "much like you or I would pick up the phone and understand the other person talking whether we had spoken to them before or not," Summerfield explained.
The system also has a number of in-built learning processes, with an underlying learning algorithm, Summerfield said.
"Speech recognition is a learning technology - you have to teach it sounds, speech patterns and accents."Finally, an error-checking process sits behind the recogniser to detect errors, and stores these in a database to allow fine-tuning of the process. "It enables the system to learn from its mistakes over a period of time," Summerfield said.
For example, if a caller was asked for a trading number, and responded: "Oh, my number is 4576", the recogniser may record the number as 04576, rather than 4576.
Specialists would then sift through these errors to correct these misunderstandings.
"In some cases, it comes down to speaker error, rather than system errors," Summerfield said.
For this reason, Syrinx has also included tutorials on the system to teach callers how to speak to the computer.