beattracking.c should do a output a float rather than an int, using a quadratic interpolation to reduce the quantisation of the tempo period candidate. all objects depending on depending on the beattracking method must be updated accordingly.

when computing the fft on a vector of zeroes, the phase elements can be found equal to zero or pi on powerpc, whereas they are always zero on x86.

this is due to the way real and imaginary parts are computed in fftw on both platforms. on x86. the result is always +0.0000, whereas on ppc, it can be -0.0000; atan2f(-0.00,0.00) = pi

see test in source:tests/python/src/spectral/fft.py, test_do_zeroes

review and amend the following lines:

• source:src/spectral/mfcc.c​ (lines 125 to 175)
• source:src/spectral/filterbank.c​ (lines 112 up to eof)

implement unit test comparing results with other implementations