In this paper, we perform an analysis of speech separation methods using deep learning. We provide a review of the literature, discussing the main features of the works based on audio and audiovisual processing, as well as the deep learning method. For the analysis of the articles, we provide a description where we identify the category, characteristics, methodology, results, and application. According to the study, we observed that the acoustic-based algorithms require audio portions of voices with external interferences to improve the intelligibility and the quality of the voice signals. Audio-visual-based methods use thousands of hours of segments of noisy videos to obtain stable performances and enhancing the quality of the separation. This latter category, also known as the cocktail party problem represents an ongoing open problem in the deep learning community.

No abstract