Skip to content

Music Source Separation with Band-Split RoPE Transformer

Summary

Current state of the art for music source separation on MUSDB18HQ (9.80 dB SDR without extra data). Introduces a band-split module that divides the frequency spectrum into subbands, each processed by hierarchical RoPE (Rotary Position Embedding) Transformers. Won the SDX23 MSS track. The band-split approach allows different frequency resolutions for different parts of the spectrum (finer in bass, coarser in treble).

Key Claims

  • Band-split processing (different resolution per frequency band) outperforms uniform STFT processing
  • RoPE Transformers in the frequency domain capture harmonic relationships better than standard attention
  • Hierarchical architecture (multiple scales) improves separation of both low and high-frequency content
  • 9.80 dB SDR on MUSDB18HQ without extra training data (0.6 dB above HTDemucs with extra data)