Fastest method of MFCC extraction on linux machine

Question

What is the fastest way of extracting mfcc from audio files in linux (Raspberry Pi in my case). I tried sphinx3 but it was slow for large files (on Raspberry Pi). SFS (speech filing system) was quite fast on windows but i could not install it on linux. Please suggest some way.

Answer 1

sphinx-train has an app to do it. I think it works reasonably hard, I've been using it on an old P4 and didn't take too long. I didn't have that much data, though.

You can use HCopy from HTK though I haven't tried it.

If you are a programmer, consider profiling the sphinx-train extractor and optimizing it. Make sure to let the authors know if you succeed.

Final note: the process does take time. It can't be avoided. Try to set-up automated scripts and leave the extraction running for the night. It's often the best way to approach it.

Answer 2

Try compiling sphinx-train or HTK with a good optimizing compiler and set the correct options for the instruction set. See, eg, raspberrypi.stackexchange.com/questions/2046/which-cpu-flags-are-suitable-for-gcc-on-raspberry-pi Looks like you have to use the right Linux distro to even have FPU support (very important!)

They don't use something like fftw or hand-coded assembly, but their implementations of FFT etc are efficient.