140630 - Chen Xue
Solved the bug caused by complex number input parameters for some wavs.
Found a bug caused by NaN input and solved it.
(The above two bugs can be called enhancements in some sense because the unordinary inputs do not exists for the sample wav.)
Refactored the codes to use float instead of double, the output is bad. Change it back to use double.
Delve into another fft library which can use double and maybe faster. Can switch to it if necessary in the future.
Use the new vocoder to train a new model.