Talking to machines: Interactive voice response gets better

Computerworld - "Touch 1 for sales, touch 2 for customer service, touch 3 for ... "

Such recorded greetings, inviting a response via the caller's touch-tone telephone keypad, are generated by interactive voice response (IVR) systems, which for two decades have been the principal communications interface between the public and corporate America, supporting self-service applications -- or at least reducing the workload on live call agents.

But these days, IVR systems are changing, leaving less and less likelihood of callers being trapped in "touch-tone hell." More corporations are switching to speech recognition so that callers are greeted by a voice that invites them to simply state their business. Reacting to the words they recognize, these systems route the calls accordingly.

Such an open-ended greeting is called a natural language system, explained Lynda Smith, division manager at Nuance Communications Inc. in Burlington, Mass., which makes the "speech engine" used in many IVRs. (Simpler, menu-structured speech interfaces are called "directed dialog" systems.)

Smith divides speech-based IVRs into four tiers. The lowest tier prompts the user to "press or say 1, and might have a "grammar" (the repertoire of words and phrases it can respond to) of 250 words. Tier 2 would be similar but with a grammar of up to 2,500 utterances. Tier 3 would add a natural language system, and Tier 4 would be capable of handling an open-ended grammar, such as would be needed for a directory look-up application. Prices range from $100,000 to $1 million, she added.

Speech recognition accuracy not an issue

Speech recognition accuracy is, oddly enough, not an issue, since the system can prompt for clarification if it's confused, explained Bob Meisel, telecommunications analyst and head of TMA Associates in Tarzana, Calif. The real issue with IVRs is containment -- how often the callers are able to complete their errands within the IVR application, without aborting the procedure by pressing 0 (or whatever it takes) to get to a live agent.

Smith said that containment rates are better for basic applications such as bank balance inquiries, but overall they can range from 40% to 90% -- assuming that a grammar has been assembled covering every way a caller might ask for the options in question. Meanwhile, the percentage of callers in touch-tone IVRs who immediately try to get to an operator by pressing zero varies from 10% to 40%, and the rate of misrouting ranges from 15% to 35%, she added. Smith claimed that the use of speech cuts the zero-out rate by up to 30% and reduces misdirected calls by up to 50%.

"Most banks were early adopters and had rates of better than 80% using touch-tone," added Mike Moors, director of sales at Genesys Telecommunications Laboratories Inc., an IVR vendor in Daly City, Calif. "A few have moved to speech and have seen a slight increase, but the rate is still in the 80s," he explained. "Health care firms usually see rates in the 15% to 20% range, since people are calling for more complicated reasons, but the system can still gather information about the caller, and 80% to 90% of the callers succeed at giving it within the IVR."