This component concatenate the cepstrum from the sequence of frames according to the window size.
It's not supposed to give high accuracy alone, but combined with LDA transform it can give the same
or even better results than conventional delta and delta-delta coefficients. The idea is that
delta-delta computation is also a matrix multiplication thus using automatically generated
with LDA/MLLT matrix we can gain better results.
The model for this feature extractor should be trained with SphinxTrain with 1s_c feature type and
with cepwin option enabled. Don't forget to set the window size accordingly.