The patent revolves around sending speech to the recognizer multiple times, and each time distorted a bit differently. All the resulting recognitions are then compared; an evaluation is done to decide which result(s) is most likely to be correct - that one is returned to the application. Fluency says they have applied this technique to all leading speech recognizers.
Via the press release:
"Dr Trevor Thomas, the inventor and Chief Scientist at Fluency, stated 'This invention will deliver important improvements to recognition accuracy and will increase the performance of our spoken dialogue systems when compared to similar dialogue systems that just make conventional use of a speech recognizer'.
Interesting stuff. We're waiting on feedback from some of our trusted sources to see just how far-reaching they think this technology is, re: transcription accuracy of continuous speech!