A neural network model for spectrogram magnitude prediction is presented. It has one convolutional layer that computes the shorttime Fourier transform. By choosing the magnitude of the spectrum as output and discarding the phase, it is possible to avoid complex number operations. The structure of the network and coefficients computation for this alternative are presented in detail. The model coefficients can be directly computed or trained with the gradient descent algorithm. In both cases, the results are satisfactory, but the obtained weights are different.
An analysis of the differences is made. The main contribution of this article is to show that the proposed model is trainable. Consequently, the coefficients can be adapted to particular problems.