How to analyze and improve the performance of a C ++ library running Android?

I want to optimize android speech processing library (arm) to make it fast enough for realtime processing. I am compiling the library already with compiler optimizations (e.g. -O3 -ffast-math -ftree-vectorize ...) but it doesn't help, so I need to optimize the code itself.

What's the best way to parse a C ++ library running on Android? Important measures can be, for example:

• Floating point operations in methods
• vectorization in methods
• ...