You can use scipy.signal tutorial and reference guide in most cases, but
there are some important differences:
* input data is assumed to be complex and two-sided spectrum is always returned
  (return_onesided argument is not implemented)
* length of FFT is always same as length of segment (nfft argument is not
  implemented)
* functions work always over last axis of array (axis argument is not
  implemented)
* if you want to have best FFT performance with pyFFTW, you should create arrays
  with empty, zeros or ones functions from SimpleSpectral instead of generic
  versions from NumPy (arrays will be byte aligned for your CPU)

WWW: https://github.com/xmikos/simplespectral