Hello everyone, thanks for your work here!

I am implementing the system for searching the keywords in input (cointinuous speech files) audio sequences (with Julius), so I have about 100 of keywords to recognize in my dictionary and I am thinking of options how to build the model to get the best results:

1. Create "2 class dictionary" containing "keywords" and "the rest" of words, or only keywords and use "-mapunk" option to classify all bad scores to it? There is no way to training the all of "outside the dictionary" words.

The second question is the training method. It will be better to add "NOISE" (sp) word between words in the training sequence?

2. i've read about short-pause segmentation. Is it possible to catch the pause between the words in continuous speech?

Thanks in advance for any suggestions or ideas what is the best option to prepare the model for this task

--- (Edited on 7/19/2013 2:03 am [GMT-0500] by Visitor) ---