This study uses wavelet decomposition (WLD) to extract mismatch negativity (MMN) from electroencephalogram (EEG) recordings. The proposed method is based on the fact that the frequency responses of WLD conform to the frequency properties of the studied signal. Thus, a suitable wavelet for the decomposition and the appropriate levels for the reconstruction required to extract the desired signal can be selected. The reverse biorthogonal wavelet of the order of 6.8 is suitable for decomposing conventionally averaged EEG recordings of the MMN activity of children elicited by an uninterrupted sound under the oddball paradigm. With seven level decomposition, the coefficients of the fifth and sixth levels are appropriate for extracting the MMN from the averaged trace. Results show that WLD can be used to extract MMN with much better properties than those of that extracted by the difference wave or an optimal band-pass digital filter. The improved properties mean that the deviant stimuli with larger magnitude deviance in contrast to the standard stimuli can elicit the MMN with a larger peak amplitude and a shorter latency. The proposed method can be used as a preprocessing step for multichannel data processing and for the study of other event-related potentials.