Dear judges,
I saw my submission having "Time" 613ms in the preliminary testing. I assume this is the average recognition time. In my tests, on an 8 core CPU, a 4096-character text should be recognized in 10ms. The maximum time for long texts is about 40ms. Very long texts are cut off to ensure this maximum.
I believe the indicated time includes model loading on each recognition. The model is loaded during the first recognition call, and loading takes about 350ms in my tests. After this, subsequent calls of the recognition function should be fast.
Please adjust the benchmark procedure to call the recognition function in a loop, without reloading the .so file on each recognition. This is how the library is supposed to be used - otherwise why build the library? If the program is to be restarted on every recognition, a simple executable file or a script would suffice.

In our final tests, this submission received the following scores (out of 100):

Code/Other Detection: 64
Language Detection: 71.1
Average Time: 586ms per sample

Final Score: 18