A Highly Scalable Music Source Separation Method based on CGAN

A considerable part of the research in the field of sound source separation is carried out on the amplitude spectrum. And the technique of separating the sound source on the amplitude spectrum seems to have encountered a bottleneck. In this paper, an audio separation method based on the amplitude spectrum, CMCTW, is proposed, which can well separate the accompaniment and vocals in music, and has better performance and effect than the cutting-edge waveform domain sound source separation algorithm. The model solves the problem of low separation quality by improving the network architecture of CGAN and proposing a loss function corresponding to the network. At the same time, a sigmoid adaptive T-F masking is proposed to improve the quality of the audio generated by the model. In the evaluation of indicators such as SDR, this method is superior to most of the current sound source separation algorithms.