Please see the IPP 7.0 release notes. The Speech Recognition domain was removed from the 7.0 release and is only part of the 6.1 product. We hope to have a KB article published that will show how to link the 6.1 SR domain with the 7.0 library, but engineering is still working on the details.
Until then, you need to either use the 6.1 library to access the SR functions or use an equivalent in the one of the other domains (such as the ippsLog functions).
The most reliable suggestion from engineeringwe have at this time is to:
create a custom library based on the version 6.1 single-threaded static speech recognition (SR) library using the the ipp_mx.h header file (i.e., %IPPROOT%\em64t\tools\staticlib\ipp_mx.h) to extract only the generic library code (use ipp_px.h for 32-bit code)
Much of the SR code relies on optimizations located in the signal processing (SP) domain, so you won't be giving up all performance advantages by using the mx layer within the SR library, since many of these SRfunctions call SP functions (which will be part of the 7.0 library).
Sorry, no KB article, yet; still working out the best solution.
If the technique you are using above results in no problem externals (that is, the only unsatisified external calls are to the other published IPP library interfaces) you should be okay.
Are functions ippsNthMaxElement_*() (e.g. ippsNthMaxElement_32s()) available somewhere else or have replacements in 7.0+?
My application is not SR-specific - that was the only function I found for calculating median on vector elements in o(n). If there is another function that can do that please let me know.
I'm using mainly IPP 7.0 and some code still needs to be compatible with 6.1.
As mentioned above, ippsr.h (and ippsNthMaxElement_*()) is available in 6.1 but not in 7.0.
What I did instead was sorting in o(N logN) - not ideal but does the job:
shift = shiftVec[len / 2];
I'm still hoping there might be a replacement for finding median or any percentile in o(N). Anyone? Any plans for this?
Thanks Sergey, read about the algorithm now and it's quite neat.
As you mention - it has a limitation based on the value range and would require dynamically allocated memory. Anyway it's an overkill for my purpose.
I hoped to find something in IPP, so that it would also take advantage of the Intel code set.
Thanks again for your help and the education...
I'm using mainly IPP 8.1 for Xcode 5.0 under OSX 10.8 and I found that the ippsr library is not included. I want to use the ippsMahDistSingle_32f64f() function. Any suggestion ? I see that if I use the 6.1 revision library I can have the speech recognition library. But is the 6.1 revision works with Xcode ?