There are three diabolical criminals on the loose, and they’re coming to steal your company’s money. But make no mistake, this is no smash and grab — they are thieves in the night, prowling cats for the chance foul up the IVR caller experience and send the call to a costly call center agent. They’re so conniving that they even have a scapegoat: speech recognition.
It happens when a speech-enabled IVR shows a rise in callers failing at prompts, resulting in transfers to an agent. “Speech is broken!” they’ll yell. The speech bandits slink off into the shadows while speech recognition gets framed.
So who are these masterminds? Just like with any great twist, they’ve been under your nose the whole time. The three usual suspects that can mask IVR trouble as a speech problem are the following:
There are uncountable parts to the deployment of an IVR: data tables, API queries and hits, client-side systems, Web services, data centers — all important to the operation of the application. And as with any system, more complexities mean more possibilities for something to go wrong.
2. Internal Marketing Miscommunication
In response to a recent promotion, callers might be saying, “Survey coupon,” but was that added to the list of possible utterances recognized? If not, then it counts as an error. Two or three of those, and just like that, they’re off to an agent.
3. User Interface Design
Is there a business rule getting in the way of a streamlined, confusion-free experience? Maybe the questions are asked in a confusing order. Do the response prompts fail to constrain possible caller responses by not giving clear instructions (e.g., please say yes or no)? Is cognitive overload being caused by too many options in one menu? All of these can cause problems with the turn-taking nature of the speech user experience, and it’ll be in the user interface design where they’ll be fixed.
So, When Is Speech to Blame?
Yes, you can blame speech recognition, but only when specific exceptions occur. These include:
- Synonyms for accepted responses are missing. For example, “billing” is an accepted response but “pay my bill” is not.
- The standard by which the speech recognizer judges utterances to be understood well enough (the “confidence threshold”) is set too high. That means the IVR is ruling out otherwise acceptable utterances.
- One of the expected responses has a weighting that is set too high. So, when the IVR compares your utterance to the list of accepted options, one of them is weighted to be chosen more frequently than the others, which can invoke an entirely different option than what the caller requested and send the caller down the wrong path.
In the world of maintaining speech applications, these are quick fixes that are usually a mere oversight.
It’s easy to see why speech recognition gets framed as the culprit so often. It’s a complex human faculty that serves as the only way the caller “touches” the IVR. As such, it’s in plain sight and is an easy target to make into a patsy. So, while speech may sometimes be the issue, due diligence to fix a problem demands that we chase down every possibility, from middleware and system interactions to marketing and user interface design, before we bang the gavel and proclaim the guilty party to be speech recognition.