This paper proposes a new music source separation (MSS) model based on an architecture with MLP-Mixer that leverages multilayer perceptrons (MLPs). Most of the recent MSS techniques are based on architectures with CNNs, RNNs, and attention-based transformers that take waveforms or complex spectrograms or both as inputs. For the growth of the research field, we believe it is important to study not only the current established methodologies but also diverse perspectives. Therefore, since the MLP-Mixer-based architecture has been reported to perform as well as or better than architectures with CNNs and transformers in the computer vision field despite the MLP's simple computation, we report a way to effectively apply such an architecture to MSS as a reusable insight. In this paper we propose a model called TFC-MLP, which is a variant of the MLP-Mixer architecture that preserves time-frequency positional relationships and mixes time, frequency, and channel dimensions separately, using complex spectrograms as input. The TFC-MLP was evaluated with source-to-distortion ratio (SDR) using the MUSDB18-HQ dataset. Experimental results showed that the proposed model can achieve competitive SDRs when compared with state-of-the-art MSS models.

If the video does not load properly please use the direct link to video