Progress report: I have been working on a custom library, to perform the same functionality as Wit.ai. For those of you who are unaware of how that service works, basically you define certain intents that you want it to trigger, and give it examples of what might be said for each, and then when you pass it a string it will send back what intent it thinks you mean.

For anyone interested, here is the approach I am using:

So, for Saera I ended up using a modification of the Levenshtein distance algorithm. To demonstrate how this works, say this is an example sentence for the Alarm intent:

[set alarm] for <time>

Green marks important words, while red marks variables - words that are expected to be replaced.

Now say the sentence we are testing is "Set an alarm at seven". In a normal Levenshtein metric, you go through and for each word, you add one to your score if you have to add, delete or swap a word, and keep it the same if the words are the same. A final score of zero means that the string you're testing is identical to the one you're testing against. But I modified it so that it scores unimportant words lower ("for" is an unimportant word in the example) - and particularly, removal of words, especially unimportant ones, is scored lower. So, going through this, we'd have:

set - stays the same - 0
an - added - 1
alarm - stays the same - 0
at - swap for "for" - 0.5 (unimportant word)
seven - swap for "time" - 1 (will be 0 when I get variables working)

Which means that the string has a score of 2.5. When run against example sentences for other events, say "[call] <contact>" for the Call intent, it scores higher - two substitutions and three additions mean that it scores a 5. Lower is better, so we take the lowest score and decide that this was meant for the Alarm intent. This is now ready to be passed to and processed by the code that was originally handling intents from Wit.ai - and this is all happening on-device!