To implement voice recognition is necessary to make a specific course of action. With a microphone turns voice recording identified and sent to the computer. The optimal reception is WAV file, since handling ease. The resulting voice recording should be divided into frames. The next step is to eliminate the undesirable effects and noises. It is necessary to record obtained at different time correspond to each other, regardless of external factors. There are many ways in which to reduce the effects of noise. To date, the most successful are the voice recognition system, using the knowledge of the hearing aid device. They are based on the fact that the ear interprets sounds not linearly but in a logarithmic scale. In view of these features is necessary to bring the frequency response for each frame of mels. This is the last step required for the subsequent conversion to vector features, which, compared to the base of voice recordings. The vector will comprise melcepstral coefficients. The resulting feature vector is added to the database for later comparison. But a more accurate alternative is to use multiple entries of the same voice. A predetermined number of voice samples may be used to train the neural network. We used learning without a teacher, because it is much more plausible model of learning in the biological system. Kohonen developed and many others, it does not need to output the target vector and therefore.