Evaluation of the Total Variation of the Spectrogram for Blind Source Separation
Classical Source Separation approaches based on the Independent Component Analysis and Sparse Component Analysis are widely used and, depending on the application, are able to achieve good results. Nevertheless, in audio signals, exploring other characteristics may lead to better performance. For example, transforming the signals into the time-frequency domain (thus obtaining the spectrogram), allows us to interpret the audio signals as images and, therefore, techniques for image procesing may be applied to bring new perspectives to the problem of source separation. In this sense, in this work, based on metrics associated with the image sharpness, we evaluate the use of the Total Variation of the Spectrogram in the problem of source separation. The preliminary results indicate that this new approach may be useful to obtain a new method for source separation.