Speech absolutely works as an interface. I wouldn't have my current job if it didn't (we make speech recognition hardware/software for mobile computing, primarily for use in industrial settings, and have been doing so for 20 years.) But to understand its real potential you have to think beyond traditional humancomputer interfaces and contexts (in short: not at a desk). What if a keyboard or even a screen is impractical? Speech can allow users to interact "hands free, eyes free" (while operating machinery, during physical activity, or while interacting with other people) This is not just about productivity, but also about safety. Cultural factors also come in to play, I believe speech is a more natural interface when it's the computer that is directing the activity (we're all used to accepting verbal commands.)

That said, you need to consider what kind of environments will it be used (noisy?) and how accurate you need the recognition to be. Part inspection in a noisy manufacturing plant will have significantly different demands than a phone based menu system. Are you going to use it infrequently or 8hrs a day? Tailoring speech recognition trade-offs is key to a successful implementation. One-size is currently still very far from fitting all.