transcription with MS Speech Recognition server API

Question

As we all know by now loading dictation grammars is not supported in the server version of MS speech API (Microsoft.Speech).

Q1: Is there a way to get around that?

Q2: One of the ideas I thought of would be to build a grammar based on a history of the conversations (chat room for instance) that people had on a particular topic. The software would need to do natural language transcription, but the subset of language used could possibly be somewhat limited.

Does this have any chances of being reasonably accurate and robust? Or is it something you are just not meant to do with non-dication grammars?

Thanks

Jan

Answer 1

I worked on a project that tried to simulate dictation transcription with a heavy use of command & control grammars. It didn't work; the uncertainty is just too high, and you get random recognitions instead of reasonable results.

Sorry.

Answer 2

I haven't worked with SAPI in a number of years, but it sounds to me like what you want is not a grammar, but a language model. This is a set of probabilities that one word follows another (my, that's an over simplification...) Anyway, most speech engines give you the ability to create a language model based on a corpus of sample text. The more text, the better it will work.