Using different sized samples for training LSTMs for speaker identification

Question

I am trying to use small chunks of voices as inputs to LSTMs for speaker Identification.
The problem is that each voice sample has a different length.

The spectrogra