Audio Source Separation and Voice Conversion, an Application in Music Industry

Voice conversion in the field of deep learning converts the voice of the source speaker into the voice of the target speaker. An application of voice conversion in the music industry has been presented in this paper. In the music industry, there is a monopoly of the music production industry that they ask different singers to sing a song, and the best version of the song is selected. This can cause degradation of morals in singers. To tackle this problem a novel unsupervised pipeline has been presented which combines Audio Source Separation (ASS) and voice conversion. For simulation, Demucs was selected as the ASS model, and randomCNN was selected as the voice conversion model. Pakistani songs were selected as the dataset and different combinations of song conversion were tested i.e., within the same gender, and different genders, etc. The result showed the conversion within the same gender and different songs over different time intervals. After evaluation, it was found that by increasing the duration of the song, content loss increases while style loss decreases.