First, there are operations between them to simulate how the ear response.

You can see from the graph under "After the IDFT (inverse Discrete Fourier Transform), the pitch information with 1/T period is transformed to a peak near T at the right side."

It is not exactly the reverse that you may expect. And

it shows that for the speech recognition purpose, only the left most part is what we care.