Music source separation extracts vocals and instruments from a mixed recording, enabling remixing, content production, and music analysis. Existing systems commonly operate on waveforms or full-band spectrograms. Waveform inputs contain many samples even in short segments, whereas a single spectrogram resolution imposes a shared time-frequency resolution trade-off across frequency ranges. We propose SubRoFormer, a Transformer-based system comprising a subband module and a residual RoFormer module. The subband module decomposes the mixture into frequency-localized, lower-rate signals for separation. It then reconstructs the target waveforms from the estimated subbands. The residual RoFormer module models temporal and cross-band relationships on temporally compressed features. It combines this context with fine-rate features through gated residual fusion to estimate source-specific complex masks. On the MUSDB18-HQ test set, training only on MUSDB18-HQ yields an average signal-to-distortion ratio (SDR) of 10.63 dB across four stems using target-specific models, achieving state-of-the-art performance among the compared systems without extra training data. With 500 additional training songs, a single model jointly separating all four stems achieves an average SDR of 11.59 dB, establishing state-of-the-art performance among the compared single-model multi-stem systems trained with extra data. Code and checkpoints are released at https://github.com/ArchiMickey/SubRoformer.
Figures & tables
Fig. 1: Overview of SubRoFormer . The fixed front end converts the analytic mixture into demodulated, decimated subbands; local complex features are processed by fine-rate boundary blocks and a temporally patched axial RoFormer; the inverse path interpolates, remodulates, and sums the estimated subbands. C , L , and K denote channels, time samples, and realized subbands, respectively.
Fig. 2: Schematic target magnitude responses of the triangular filter bank, with illustrative frequency spacing. The highlighted interior band Ak(f) has lower edge fk,ℓ , peak fk,p , and upper edge fk,u . Neighboring responses overlap along their slopes; the first and last bands are one-sided. The dashed line shows their unit sum over 0≤f≤Fs/2 . These are ideal targets, rather than the realized finite-length FIR responses.
Fig. 3: Per-band analysis and rate conversion: (a) the input x[n] ; (b) the band-pass filtered signal xk[n] ; (c) the analytic signal xka[n] ; (d) the complex baseband signal bk[n] ; and (e) the decimated sequence zk[m] . Uppercase symbols denote the corresponding spectra. The example uses Fs=48 kHz, fk=10.4 kHz, approximately 800-Hz band support, and D=60 , giving a reduced sampling rate of Fs/D=800 Hz.
Reconstruction SDR (dB) ↑
Transform
Rate
Mixture
Vocals
Bass
Drums
Other
LSD (dB) ↓
NME ↓
STFT [ 19 ]
4.271
138.58
138.47
138.75
138.64
138.63
0.0000
0.0000
SubRoFormer mel–linear
4.267
28.04
26.80
47.78
26.58
29.03
3.3375
0.0264
SubRoFormer ERB–linear
4.267
28.04
26.80
48.26
26.56
29.11
3.3318
0.0265
Auditory ERB [ 46 ]
0.125
0.54
0.05
1.00
0.88
0.15
113.0755
0.9353
PQMF [ 18 ]
1.000
11.30
10.25
13.41
12.03
10.38
3.5195
0.2770
TABLE I: Transform reconstruction quality on the MUSDB18-HQ test set. SDR is delay-aligned and summarized by the median over 50 songs per signal; LSD and NME are medians over 250 song–signal pairs. Rate counts real scalar coefficients per sample per channel, including padding. Higher SDR and lower LSD and NME indicate better fidelity.
Model
Params.
N
Voc. ↑
Bass ↑
Dr. ↑
Oth. ↑
Avg. ↑
Single-stem models
BSRNN [ 14 ]
37.6M
4
10.01
7.22
9.01
6.70
8.24
BS-RoFormer [ 15 , 35 ]
82.8M
4
11.02
11.58
9.66
7.80
10.02
Mel-RoFormer [ 35 ]
94.8M
4
11.60
–
9.34
7.93
–
SubRoFormer-S (Ours)
22.52M
4
11.34
9.18
10.60
8.16
9.82
SubRoFormer-B (Ours)
97.17M
4
12.63
10.20
10.40
9.27
10.63
TABLE II: MUSDB18-HQ metrics for systems trained without extra data. Metrics for research baselines are taken from their original papers.
Fig. 4: Mean test SDR (dB; higher is better) on MUSDB18-HQ versus total deployed parameters for selected systems trained without extra data. Single-stem systems, including both SubRoFormer sizes, use four networks; multi-stem baselines use one. The plotted BS-RoFormer and Mel-RoFormer points are six-layer configurations. Baseline results are drawn from [ 11 , 14 , 15 , 35 , 13 , 62 , 36 ] .
Total (M)
Representation + mask (M)
Separator (M)
BS-RoFormer L6
72.19
50.87 (70.47%)
21.32 (29.53%)
Mel-RoFormer L6
84.19
62.87 (74.68%)
21.32 (25.32%)
SubRoFormer L4+2
24.08
2.76 (11.45%)
21.32 (88.55%)
TABLE III: Parameter distribution of single-output variants with matched separators.
Variant
SDR ↑
Δ SDR ↑
SubRoFormer-S (400k updates)
10.88
0.00
BS-RoFormer L6 ∗ [ 15 ]
10.18
–
Mel-RoFormer L6 ∗ [ 35 ]
10.08
–
A1: Front-end representation
Global-STFT band-split
9.43
−1.45
Mel band-split
8.88
−2.00
TABLE IV: Vocals-only ablation results on MUSDB18-HQ.
Model
Params.
N
Extra
Voc. ↑
Bass ↑
Dr. ↑
Oth. ↑
Avg. ↑
Single-stem models
HT Demucs (FT) [ 13 ]
41.4M
4
800
9.20
10.39
10.08
6.32
9.00
BSRNN [ 14 ]
37.6M
4
1750
10.47
8.16
10.15
7.08
8.97
BS-RoFormer L12 [ 15 ]
93.4M
3
450
12.72
13.32
12.91
9.01
11.99
Multi-stem models
Hybrid Demucs [ 12 , 13 ]
83.64M
1
800
8.75
9.13
9.31
6.18
8.34
TABLE V: MUSDB18-HQ metrics for systems trained with extra data and a contextual system with unknown training data. Metrics for research baselines are taken from their original papers.