rss
 
comment(s)

archives
J|F|M|A|M|J|J|A|S|O|N|D
(20##) 10 9 8 7 6 5 4 3 2 1 0 <
 
DesktopWeb FormText   /eva : dictation resource kit (round 2)Tue, 01 Sep 2009 23:28:04 GMT # 

integrated DRK grammars into /eva. first, the code works on both Vista and Windows 7. i'm still trying to figure out why the previous n-gram loading code worked on Vista, but not Windows 7. second, this allows me to automate generation of the grammars nightly on a deployment machine. the only requirement is installing the DRK to the deployed machine, which is much easier to swallow than installing Visual Studio and Speech Server 2007 to get the Conversational Grammar Builder installed. that said, i'd still rather have a simple executable that i could hand a corpus text file and get back a compiled n-gram (.cfg).

i've only done minimal runtime testing, but so far, the results from using an n-gram with AppendRuleReference seem to be better than using a DRK language model with AppendDictation(topic). that's strictly from looking at the SemanticValues returned from Recognition. second, i'd also have to rework my search logic. the problem is i normalize the music library in my own code and store the original values and normalized values to a dataset. but then the DRK can normalize the values again, so i'd have to remap that normalized value back to the dataset to get the search logic to find more matches. e.g. my logic currently normalizes '808 state' as it is, but the DRK will normalize to 'eight zero eight state'. for now, i'm just skipping the step of using the DRKs normalizer.

so i've still got problems. the n-gram seems to give better results but there isnt a way to automate generating grammars and the code isnt working on Windows 7. vs the DRK language models, which dont seem to perform as well recognition-wise for this task, but they can be generated automatically on a deployment machine and the code is working on Windows 7. er, um ... so i'm stuck with using the DRK for now.