These things do not work on an American accent, only on one from a specific part of the country.
I'll call bullshit. I work daily with a voice recognition system integrated with Word and MS-SQL to process medical transcription. (Not by choice, mind you, but it's all there is in this field. Oh, and patents. Lots of patents.) It seems to do well with a good variety of accents, many of them ESL speakers with funny accents and mispronunciations. Trust me, it all still needs an editor to perfect the output, but it's far more capable of getting readable output on a page than you think. If it can do that, then it's good enough for natural language machine-human interaction.