Abstarct: Separation of sing voice from music has always attracted much attention in the fields of audio processing in recent years and is very useful in many applications such as singer, music information retrieval, automatic singing transcxription and content-baxsed music retrieval. Because speech signal is a nonstationary signal ,the time frequency analysis is one of the best method to extract features from speech signal. In this regard, in this thesis, we presented a method baxsed on time-frequency analysis using CNN and FCNN deep neural networks to separate voice and music. The simulation results on the MIREX2018 databaxse showed that the short time fourier transform method produced 85.02% accuracy rate for CNN and 87.18% for FCNN while the wignerville method produced 86.15% for CNN and 87.03 for FCNN and in the Gabor method prodused 82.74% for CNN and 85.22% for FCNN. Morever the experimental results showed that in all methods except one mode of fully connected convolutional neural network (FCNN) method have better results and generally provides better performance.