We have been developing CeedVocal SDK since 2008, it's based on Julius & FLite open source projects.
Here's some context: we wanted to make our app (Vocalia) for speech recognition back in 2008 and basically picked Julius (hesitated with Pocket Sphinx, which appears to be good as well) and optimized its file format so that it would boot in 1-2 sec instead of 20sec on the original iPhone. Then we dutifully trained our own acoustic models in 6 languages. We designed the API, and eventually decided to offer it to other developers as an SDK.
CeedVocal basically supports 2 modes of operation:
- matching of words (or small phrases)
- keyword spotting
In the first mode of operation, it tries to align the input speech to a word (or phrase) in its list of acceptable input. This forces the input to a pre-known word, even if the speech is something else. Accuracy is good. In the second mode of operation, it will try to pick one of its keywords into the stream of speech. This is a difficult case, and it can be less accurate.