A Band-Split State Space Model for Accordion Left- and Right-Hand Source Separation With Pedagogical Applications

Abstract

“Hands-separated practice followed by hands-together integration” is a long-established pedagogical paradigm for the accordion and other keyboard instruments. During self-practice, however, this paradigm relies almost entirely on printed scores or pre-prepared hand-separated audio tracks, and it is therefore difficult to apply directly to real performance recordings or to pieces for which no score is available. To address the task of accordion left- and right-hand source separation, which to the best of our knowledge has not been specifically studied, we propose an end-to-end deep learning model called Accordion Music Source Separation with Mamba (AMSSM). AMSSM encodes the spectral structure of the accordion through a non-uniform sub-band partition that is dense at low frequencies and sparse at high frequencies. It builds a unified, linear-complexity sequence backbone using bidirectional Mamba-2 along the time, frequency, and channel axes. It then uses a dual-branch separation head, in which a mask branch and a map branch operate in parallel to carry the additive mixing prior and the residual refinement, respectively. We also construct a hand-separated accordion dataset for this task, comprising 542 pieces and about 4.5 hours of music content. Experiments show that AMSSM attains an overall signal-to-distortion ratio (SDR) of 23.97 dB on the test set. This separation quality is substantially higher than that of representative open-source music source separation (MSS) methods and allows the model to reliably recover perceptually clear left- and right-hand voices from hands-together recordings. This capability is of potential value to hands-separated practice, a pedagogical step that has long been under-served by existing tools. We further discuss the technical readiness of AMSSM as a candidate hands-separated practice aid along two dimensions: alignment with objective task requirements, and protocol-level recommendations for future pedagogical evaluation. The empirical pedagogical effectiveness of AMSSM is not claimed in this paper and is left to subsequent controlled studies on actual learners. At a conceptual level, the proposed approach can also be extended to other instrument-internal separation tasks with similar voice-independence properties, such as piano left- and right-hand separation and pipe-organ multi-voice separation. The source code will be released upon publication of this paper at https://huggingface.co/spaces/wuchenaudio/ams