Avoid the Phone Slam with a Finely Tuned IVR

Overtesting can also be a problem. Tuning brings diminishing returns. In development, the first few major revisions often yield significant improvements, with later revisions producing smaller gains. So the question becomes, when should the company stop tuning its system? Tuning the first release often yields significant improvements (as much as 25 percent in some cases), so businesses go through a few tests, according to LumenVox’s Miller. Smaller contact centers are less likely to conduct more tests and shoot for smaller incremental (say a few percentage points) gains than larger organizations, which are better able to cost justify such investments.

Building a Library

In addition, tuning still mixes art with science. IVR systems are based on language, which is imprecise and open to interpretation. Consequently, building a comprehensive speech library is an important part of speech recognition IVR applications. Here, a business develops a collection of phrases that are meaningful to the corporation, customers, and business units; after identifying the phrases, the business monitors their use to determine how well the system performs. Generating a large number of hits will only be useful if the firm has the extensive resources to review, interpret, and analyze them all. When building the library, an organization needs to avoid phrases that don’t provide much value or are so rare that they don’t merit inclusion because there is a significant cost in analyzing them.

The enterprise then needs to check the words in context. An organization can select seemingly great phrases, but how customers use them is often unpredictable, and in some cases, the phrases fool the system. “Cancel” and “can sell” can sound the same, for example, so the ­organization may need to find a substitute for that term.

How Much Confidence Do You Need?

Voice recognition confidence levels are another area of constant adjusting. A confidence level (for example, 70 percent) reflects the degree of certainty required for the speech engine to register the word as spoken. A trade-off exists in setting confidence levels. Low confidence levels generate more false positives, instances when the speech engine mistakenly identifies the wrong phrase. For example, the engine thinks it heard “more dates” when, actually, “more data” was spoken. Corporations typically use low confidence levels when they absolutely do not want to miss certain phrases, such as “I will sue you.” The trade-off is that they then must wade through a potentially large number of false hits.

Higher confidence levels result in more misses, instances when the speech engine fails to correctly identify an actual spoken phrase. High confidence levels are used when a business wants a general overview of word usage rather than a pinpoint analysis. The business trades off less work for imprecision. Because of the complexity of the various factors, companies need to budget time, money, and effort into adjusting their speech library’s confidence levels.

Time for a Change?

However, tuning is an ongoing issue. Many companies put an IVR script in and leave it running unchanged for months, years, and even decades. The reality is customer interactions are dynamic, so firms need to check how well their systems are operating. There needs to be follow-up once the system is installed. A critical step in any well-conceived speech system is a periodic content audit: a manual review of a sizable sampling of call recordings. Listening to actual calls from start to finish is a tedious process, but one that allows the business to tune its speech library by capturing and examining the actual utterances used in customer interactions.