Menus are all about capturing the caller's intent. There are many ways to do this: traditional multi-item menu, yes/no question, two-item menu, SLM (open-ended prompts supported by a statistical language model), plus a decision between speech and DTMF. Determining which approach is right isn't always straightforward. In the sections on each of these much more detailed information is given, but here is an overview.

If you have more than two items, your choices are an SLM or a multi-item menu. The multi-item menu is probably the most common structure for starting an IVR system, what people think of when they think of an IVR or a phone tree. But is there a point at which a standard menu is too big and an SLM should be used? Yes. Defining that limit is difficult. See this section for more information.

A third situation arises with complex questions and choices. Since the goal is to get callers through the system, you're better off asking more questions that are shorter and simpler than trying to combine things into fewer questions. Perceived efficiency is the goal. A key driver of a caller's perception of efficiency is how easily they get through a task. Plus, if a question is complex enough, they might not get through at all.

There are two slightly different applications of this principle. The more straightforward is simply gathering a single piece of information at a time rather than trying to get more than one with a multi-slot grammar. This situation isn't even a menu. The second situation is a little different. It's where presenting everything at once is just too complex to sort out effectively in the caller's mind. A series of shorter questions is the way to go.

Consider a health care application where the caller has selected to hear information about preventative care. There are three different kinds: baby wellness visits, well woman checkups, and general physicals. Trying to label these clearly enough to put into a menu proves to be rather difficult (never mind the fact that nobody wants to say “well woman checkup” out loud, especially in a non-private situation.) The designers in this case settled on a series of yes/no questions about the patient's age and gender. These are very easy to answer and puts the onus of figuring out which type of wellness visit the caller wants onto the system rather than the caller.

When you have two items, it can either be a menu with just those two items, or it can be structured as a yes/no where one of the two options is presented. Which approach to take can be one of the harder decisions to make. The section on yes/no vs. two-item menus goes into a lot of detail.

SLMs are most appropriate where there a large number of options (as in, a couple dozen or more), and lots of utterance data to build the model from.

Speech vs. DTMF is usually more of a question about technology and user environment than the particular menu, but there are some exceptions. Within a speech application, there may be a point where a DTMF selection is necessary. If there is a large group of options from which the menu for a given caller will be created dynamically, it might not be feasible to have grammars built that contain all of those options.