No matter how carefully you might have crafted your VUI design, or how diligently the design was implemented, or how thoroughly the implementation was tested, your application will need regular and careful tuning once deployed if your aim is to maintain a world-class, highly usable voice solution.
 
To effectively tune your application, you should have at your disposal three sources of information:

1. Call Logs: which will enable you to identify patterns across calls (e.g., where are people hanging up);
2. Call Recordings: which will enable you to understand the nature of a problem (why are people hanging up?); and
3. Your callers: usually, you do this by assessing their level of satisfaction with the solution. 

Here are the basic questions that need to be asked in order to begin tuning a voice application:
 
Where are people hanging up? A hang up prior to completion of a task is usually a sign of frustration. If the goal of your application is automation, your first tuning task is to identify such hang up spots in your application and understand why people are hanging up.
 
Where are people asking to be routed to an agent? If you have designed your application with the goal of empowering the caller, you must have provided the caller with the option to route to an agent. A caller actively asking to speak to an agent is a caller who has decided that the application is not successfully enabling them to serve themselves. This is especially true of callers who have engaged the application over several minutes of interaction and then decided to bail out.
 
Where are people saying the wrong things? The aim here is to identify those spots in your application where no-match failures are significantly higher than the average or the expected. The remedy is to listen to the prompt the caller hears and then listen to what people are saying in response to that prompt. In such situations, adjust your application by either rewriting the prompt or by adding to the language with which system is listening to how callers are responding.
 
Where are people not saying anything? These are the spots in your application where the caller goes quiet on you. This occurs usually because the prompt is confusing or the caller was asked for some information that he doesn’t have (or doesn’t have ready access to, such as a subscription ID or an account number). If the issue is with lack of clarity of ambiguity, then recraft your prompt. If the issue is with lack of readiness, then provide the caller with the time he needs to retrieve the information you need from him or suggest that he calls back when he has the information handy. Another strategy is to inform the caller at the very outset of the interaction that the subscription ID or the account number will be needed.
 
Where are people speaking too soon? At times, callers are impatient and speak sooner than they should, often missing crucial information or instructions. To remedy, either turn the barge-in setting off, or recraft the wording of the prompt the caller is interrupting.
 
What type of noise level sare your callers calling from? When you listen to your recordings, pay attention to the noise level and how the noise is affecting the no-match error rates.
 
What options are people asking for? If you discover that 80 percent of your callers are checking their savings balance, then ask 100 percent of your callers if they are calling about checking their Savings balance. By definition, 80 percent of the time you will be right.
 
How are people feeling about the application? You can probably get a good sense of how people feel about the application by just listening to the tone of their voices in your call recordings. 

Ahmed Bouzid, Ph.D., is head of the product group at Angel.com and co-author of The Elements of VUI Style. He has more than 15 years of experience in the speech and natural language processing. He can be reached at bouzid@angel.com.