Benchmarking Music Demixing Models for Deep Drum Source Separation

Traditionally, drum source separation has been tack-led using nonnegative spectro-temporal factorization methods. Only recently, deep learning showed unprecedented performance in separating five stems from a drum mixture, namely, kick drum, snare drum, toms, hi-hat, and cymbals. The literature, however, still lacks a thorough comparison of the techniques readily available in the context of music source separation. In this paper, we conduct a first benchmarking analysis of music demixing models tailored for deep drum source separation. We evaluate a range of state-of-the-art neural network architectures, including HT-Demucs, MDX23C, and BS-RoFormer, trained using StemGMD, a large-scale dataset of isolated single-instrument drum stems. Besides demonstrating that said architectures out-perform the state-of-the-art method for drum source separation, we discuss their strengths and weaknesses, giving insights into their performance and ultimately offering valuable guidance for researchers and practitioners willing to develop drum demixing models for different applications, among which those related to music making, personalized listening, and online music education stand out.