Donnerstag, 6. August 2009

sam

I already mentioned it in the last post: A new application has been added to the simon application suite: sam.

sam is targeted towards power users who want to tweak and improve their acoustic model manually to improve recognition rates even further.

sam will include a sophisticated testing framework to immediatly receive feedback on changes in the model configuration. In fact during optimizing models manually, I realized that IMHO a well working, automated model testing framework is the most essential part in manual optimization as it makes the impact of changes immediatly visisble.

In contrast to simon, sam will not hide any of the internal workings from the user (due to the different target group) so the logs of both the building and the testing of the model are displayed and the whole operation can be double-checked for errors or warnings.

An initial, working version is already available through SVN.

Selecting the input files:


Building the model:


Testing the model:


Test results:


As you can see, simon will run the recognition with the generated models on the trainingssamples to see if simon correctly recognizes their contents. The algorithm already recognizes and considers confidence scores of the recognition results which is why in the screenshot you can see the recognition rate of e.g. "NULL" not being 100% even tough every instance of it was recognized correctly (5/5).

Btw: This is a well trained, rather small model which really works very well in practice so don't be alarmed by the very high recognition rate...

Greetings,
Peter

Keine Kommentare: