Leveraging Information To Increase Call Completion In Speech Recognition Applications

The goal of a speech recognition Interactive Voice Response (IVR) system is to successfully provide and/or capture automated data through to call completion. It can be generally assumed, that a caller that quickly and efficiently navigates the IVR without being asked to repeat or confirm data is a satisfied user of the system. Conversely, a caller that is prompted several times to repeat a spoken response because the system is not confident in the utterance, can become frustrated, displeased with the system and ultimately the company, as well.

In this paper, we will introduce a method of speech recognition application design that will help to alleviate the need to perform multiple confirmations of a caller and improve call completion rates. By leveraging even one more piece of additional data for cross-reference, we increase completion percentages and improve a caller's experience.

Case Example

For this analysis we will use a typical, pharmacy speech recognition IVR application that asks for a caller’s prescription number and their 10-digit phone number to refill their prescription.

The Challenge

For a given string of numbers, such as prescription number or phone number, there is a small chance for each of the digits to be recognized by the system incorrectly. A single spoken digit is recognized with 98% accuracy. Across a 10-digit string we might see an 18% chance that one number is incorrect, or 82% recognition rate. Therefore, if you take 2, 10-digit strings, per our example, you might have a 33% chance, or 67% recognition rate, that one number will be wrong.

First, the caller speaks their 10-digit prescription number. In our case, the system recognized only the first 9 digits of the prescription number, or “123456789?.”

	Requested Information		Recognized Digits
	Prescription Number		123456789?

While we weren’t confident enough to confirm a perfect match, there was enough information to reduce our data set and discover there are three customers that match that profile: Ann Moss, Freddie Young, and Ted Gonzalez.

Callers	Prescription#	Phone	SSN
Ann Moss	1234567899	8587070209	7645
Freddie Young	1234567898	8587070707	7777
Ted Gonzalez	1234567897	8587070307	7245

Next, we ask for their phone number.

	Requested Information		Recognized Digits
	Prescription Number		8587070?07

Again, there is one digit that is unclear in the caller’s utterance, however we were able to remove Ann Moss from our list of contenders, and can focus on Freddie Young and Ted Gonzalez.

Callers	Prescription#	Phone#	SSN
Freddie Young	1234567898	8587070707	7777
Ted Gonzalez	1234567897	8587070307	7245

Next, instead of using the confirmation method and asking them to repeat the numbers again, we ask for the last 4 digits of their SSN. Once we have this 3rd piece of information, we can cross-reference with our two remaining customers and find the correct match. In this case, it is Freddie Young.

	Requested Information		Recognized Digits
	Prescription Number		123456789?

The end result is that we have a three-interaction transaction toward call completion, which did not require any confirmations. This is considered a successful customer call, and leaves Freddie Young satisfied with the experience of using a speech recognition IVR system.

Conclusion

Our prescription refill example showed us that leveraging additional pieces (or just one piece!) of information could not only increase call completion rates, but also improve a caller’s experience. By simply using our database information cache effectively we reduced caller frustration and avoided confirmations altogether.

For answers on building speech applications or for further information how LumenVox technology can help you be a successful speech solution designer, please go to www.LumenVox.com, or call us toll free at (877) 977-0707. We’re here to help.

Leveraging Information In Speech Recognition Applications

Print Version

View PDF file
View a print ready version of this page.

Key Takeaways

The goal of this white paper is to:

Provide a more accurate picture of application performance statistics and how to use them in the tuning process.
Show how to reduce the number of calls needing lengthy analysis.
Demonstrate how to save time, resources and money.