The answer is never. The reason is that different kinds of UIs are effective for different kinds of tasks.
I can't remember who I heard this from, but they asked the audience to imagine a speech-based steering wheel for cars. "Turn left... more... more... MORE!!! LESS! LESS!"
Speech also has problems with sensitive data in public places ("Please enter in your password" or "Your stock portfolio is now worth ..."), has problems with noise, and can also interfere with some parts of your cognitive processes. In addition, speech is serial, which means that you can pretty much only process one stream at a time. Contrast this with a web page that you can skim, or a visualization that can show a lot of data.
Different kinds of interactions have different tradeoffs. Speech is good for some tasks, and not very good for others.