As a general rule, smoke and mirrors anywhere there are rough edges. In your case, if your demo does not work under less-than-ideal speech recognition conditions, I would not let them speak to the prototype. You could ask the audience for queries and speak them yourself.