Automatic speech recognition is performed by applying the Maximum a-posteriori decision rule to estimated probability density functions (pdfs) of the signals from the different words in the vocabulary. These pdfs are assumed hidden Markov models and their parameters are estimated from labeled training data. Normally, modeling is applied to the inverse Fourier transform of the log-spectrum which is known as the Cepstrum. In this talk I shall review the principles of automatic speech recognition and some interesting properties of Cepstrum. These properties are used to improve robustness of the speech recognition system when the speech is corrupted by noise.