Having done some work with watermarking, I would assume it's extremely accurate. There is so much information that makes it through the transmission process: your speech cadence, inflection, pronunciations, word choice, interjections, background noise, etc. The only way I'd feel reasonably comfortable defeating something like voice printing is using a text-to-speech engine.

You are thinking about it from someone trying to understand another person through distorted comms. You aren't thinking about every other signal in the "speech" that is really quite easily identifiable with a reasonable margin of error.