MLP possesses 26 input nodes for the 26 Mel-scale coefficients, which in this instance are averaged over a number of segments, to covers the whole words. The two words' speech signals are converted into 26-element vectors.