Bottom line: speech recognition in the general case (more than a few predetermined words) is only as good as the 1) acoustic model (which utterances were heard), and 2) language model (how do we group the utterances into words).

This requires massive amounts of labeled data. This is why Nuance is king and few others come close - the amount of labeled data necessary to catch up is astounding. Not to mention a patent minefield to navigate.

This is unfortunately one field in which open-source alternatives face real obstacles and won't be viable in the near future.