SCISSOR

Audio examples

SCISSOR: Score-Conditioned Instrument Source Separation for Orchestral Recordings

Abstract

Orchestral separation recovers instrument sections from mixtures in which shared pitches, harmonics, and timbres obscure source identity. An aligned score provides instrument labels, note pitches, and activity times. A score-informed approach appends piano rolls to audio features before mask prediction. We introduce SCISSOR (Score-Conditioned Instrument Source Separation for Orchestral Recordings), which uses the score to form a frame-wise query for each source. Each query matches a shared audio representation, and a softmax over instrument and background slots jointly assigns overlapping time–frequency evidence. The queries retain instrument identity even when notes are missing from the score.

After training on SynthSOD and a small set of URMP and PHENICX-Anechoic recordings, SCISSOR achieves the highest average SDR on held-out real recordings. With SynthSOD-only training, it leads on SynthSOD and zero-shot PHENICX-Anechoic, and improves on its audio-only control on zero-shot URMP. SCISSOR also degrades less under score corruption than the evaluated score-based baselines.

SCISSOR architecture. The audio pathway encodes the mixture spectrogram into 96-band features; the score pathway turns aligned note events into frame-wise source queries with learnable slot identities. Cosine matching and a five-slot softmax assign audio evidence to sources, and a shared FiLM complex refiner produces masks for the four string stems.
Figure 1SCISSOR overview
6.80 dBHeld-out URMP, trained with real recordings
6.31 dBSynthSOD test set, 59 pieces
1.85 dBZero-shot PHENICX-Anechoic

Average museval SDR over violin, viola, cello, and double bass; the highest among the evaluated systems in each setting.

How to listen. Each box is one full recording. Columns are systems, rows are the mixture and the four string parts. Switching players within a box keeps the playback position, so you can compare systems at the same moment. On narrow screens, scroll each table sideways.