Learning Where to Split: Dynamic Band Splitting for Audio Source Separation
Authors: Bingzhe Zhang, Kai Li, Xiaolin Hu
Organizations: Department of Computer Science and Technology, Institute for AI, BNRist, Tsinghua University, Beijing 100084, China · IDG/McGovern Institute for Brain Research, Tsinghua University, Beijing 100084, China · Chinese Institute for Brain Research (CIBR), Beijing 100010, China
Audio source separation models must preserve fine spectral structure while keeping the cost of the separator tractable. Band-split architectures reduce this cost by manually partitioning the spectrum into predefined frequency bands and encoding each band as a token, but the same boundaries are used for every mixture. Because informative spectral structure varies with source content, fixed boundaries can merge discriminative components or spend tokens where little resolution is needed. We introduce Dynamic Band Splitting (DBS), an end-to-end framework that adapts boundary locations and band widths to each mixture under a fixed band-token budget. DBS predicts content-conditioned boundary scores within frequency blocks, selects a prescribed number of boundaries using hard Top-K, and applies content-dependent pooling to obtain compact band tokens. After separation, it smooths and expands these tokens to the original STFT grid, refines them along frequency, and decodes complex masks with a frequency-shared head. DBS improved separation over the corresponding static models on DnR v2, MUSDB18-HQ, and EchoSet, with DBS-BSRNN achieving relative gain of 29.0% in average SI-SDR on DnR v2. These results and boundary analyses on DnR v2 show that content-adaptive frequency partitioning benefits multiple separator backbones under a fixed band-token budget. Code and audio demos are available at https://cslikai.cn/Dynamic-Band-Splitting
Figures & tables
Figure 1: DBS partitions across a 30-second Tears of Steel excerpt. Each column represents a non-overlapping three-second window, showing its midpoint frame, dominant acoustic content, and predicted partition. Colored lines highlight five learned boundaries per block; faint lines show the others, and black lines mark fixed block limits.
Figure 2: Overview of DBS, using cinematic audio separation on the DnR v2 dataset as an example.
Figure 3: Detailed dynamic band splitting within block r .
Figure 4: Full-resolution reconstruction within block r .
Method
Music ↑
Speech ↑
SFX ↑
Avg. ↑
Params (M) ↓
MACs (G/s) ↓
Conv-TasNet ( Luo and Mesgarani, 2019 )
0.30
8.50
2.00
3.60
5.3
19.8
MRX ( Petermann et al., 2022 )
4.20
12.30
5.70
7.40
30.5
10.6
BSRNN ( Luo and Yu, 2023 )
5.50
13.80
1.80
7.03
52.8
18.2
TIGER ( Xu et al., 2025 )
5.14
13.51
3.66
7.44
1.4
10.8
BSRNN (128) † ( Luo and Yu, 2023 )
4.83
14.87
3.38
7.69
35.5
76.6
DBS-BSRNN (128) †
6.48 +1.65
15.01 +0.14
5.73 +2.35
9.07 +1.38
4.7
124.0
Table 1: Source separation performance on the DnR v2 test set, measured by SI-SDR (dB). Models marked with † are trained from scratch with 128 bands. Red subscripts indicate improvements over the corresponding static baselines. MACs are reported per second of input audio.
Model
Voc. ↑
Drums ↑
Bass ↑
Other ↑
Avg. ↑
Params (M) ↓
MACs (G/s) ↓
BSRNN † ( Luo and Yu, 2023 )
6.01
6.23
4.49
3.62
5.09
71.7
21.8
DBS-BSRNN †
6.70 +0.69
6.44 +0.21
4.58 +0.09
4.01 +0.39
5.43 +0.35
21.7
20.8
BS-RoFormer † ( Lu et al., 2024 )
5.14
5.76
3.95
3.17
4.50
378.5
292.7
DBS-BS-RoFormer †
7.64 +2.50
7.72 +1.96
5.29 +1.34
4.72 +1.55
6.34 +1.84
33.9
290.5
MelRoFormer † ( Wang et al., 2023a )
9.20
8.29
6.05
5.48
7.25
98.7
171.7
DBS-MelRoFormer †
10.02 +0.82
9.30 +1.01
6.96 +0.91
5.98 +0.50
8.07 +0.82
19.2
252.9
Table 2: Music source separation performance on the MUSDB18-HQ test set, measured by SDR (dB). All models ( † ) are trained in this work without external data and jointly separate the four stems.
Variant
Music ↑
Speech ↑
SFX ↑
Avg. ↑
BSRNN (128) ( Luo and Yu, 2023 )
4.83
14.87
3.38
7.69
Uniform
5.58
14.43
3.99
8.00
Mel bands ( Wang et al., 2023a )
2.85
14.65
0.10
5.87
Learnable filterbank ( Dai et al., 2023 )
-11.47
8.62
-32.30
-11.72
DBS-BSRNN
6.48
15.01
5.73
9.07
Table 3: Comparison of frequency partitions and front ends on the DnR v2 test set using a BSRNN separator and 128 bands.
Figure 5: Boundary adaptation on DnR v2 under a fixed 128-band budget. Replacement is the percentage of internal boundaries not shared by two partitions of the same block. Crosses mark repeated inference on identical inputs. Center lines mark medians, boxes span interquartile ranges, and whiskers extend to the most extreme observations within 1.5 interquartile ranges of the quartiles.
Variant
Music ↑
Speech ↑
SFX ↑
Avg. ↑
1Block
3.73
13.39
1.77
6.30
2Block
5.72
14.52
4.71
8.32
Full DBS (4 blocks)
6.48
15.01
5.73
9.07
Table 4: Effect of the number of frequency blocks on DBS-BSRNN performance on the DnR v2 test set.
Figure 6: Boundary probabilities p(f) and selected boundaries for the 1Block variant on a DnR v2 test window. Most selected boundaries fall at low frequencies, leaving the upper frequency range sparsely partitioned.
Variant
Music ↑
Speech ↑
SFX ↑
Avg. ↑
Full DBS
6.48
15.01
5.73
9.07
Magnitude-only predictor
4.33
14.02
2.43
6.93
w/o explicit energy
5.88
14.50
5.11
8.50
w/o acoustic transition prior
0.49
12.83
−1.04
4.09
w/o confidence path
5.06
13.90
3.62
7.53
w/o frequency refinement
0.11
13.26
−1.61
3.92
Table 5: Boundary prediction and frequency reconstruction ablations for DBS-BSRNN on the DnR v2 test set. Magnitude-only predicts boundaries from magnitude statistics without learned spectral features. Rows marked “w/o” remove the named component from full DBS.
Appendix figures & tables12 assets
Supplementary material from the paper’s appendix.
Appendix
Dataset
Rate (kHz)
Duration (s)
Train
Test
DnR v2
44.1
60
3,406
973
MUSDB18-HQ
44.1
variable
100
50
EchoSet
16
6
20,268
2,650
Appendix
Table 6: Dataset sampling rates, recording durations, and official training/test partition sizes. Duration refers to a complete mixture or music track.
Dataset
STFT bins
Blocks
Band tokens
DnR v2
1025
4
128
MUSDB18-HQ
1025
4
128
EchoSet
321
4
67
Appendix
Table 7: Dataset-level STFT resolution and band-token budget shared across DBS backbones.
Backbone
Original band interface
Retained separator
BSRNN
Fixed non-overlapping bands; per-band input and mask heads
Temporal and inter-band RNNs
TIGER
Fixed band splitter and restoration head
FFI blocks
BS-RoFormer
Fixed non-overlapping bands; multi-band masks
Hierarchical RoPE Transformers
MelRoFormer
Overlapping Mel bands; Mel-band masks
Hierarchical RoPE Transformers
Appendix
Table 8: Integration of DBS with four band-split separators. DBS replaced each model’s band interface and mask reconstruction while retaining its separator core.
Method
SDRi ↑
SI-SDRi ↑
Params (M) ↓
MACs (G/s) ↓
A-FRCNN-16 ( Hu et al., 2021 )
9.64
8.76
6.1
81.3
TDANet Large ( Li et al., 2023 )
10.14
9.21
2.3
9.2
TF-GridNet ( Wang et al., 2023c )
13.73
12.85
14.4
323.8
BSRNN ( Luo and Yu, 2023 )
12.75
12.23
26.0
98.7
DBS-BSRNN
13.21 +0.46
12.66 +0.43
7.1
38.1
TIGER ( Xu et al., 2025 )
14.22
13.73
0.8
15.3
Appendix
Table 9: SDRi and SI-SDRi (dB), parameter counts, and MACs per second of input audio on EchoSet. Gray rows mark our DBS results; red subscripts show gains (dB) over the paired baseline.
Boundary source
Music ↑
Speech ↑
SFX ↑
Avg. ↑
Uniform
5.07
13.31
5.02
7.80
BSRNN layout
6.04
14.36
5.46
8.62
Learned DBS
6.48
15.01
5.73
9.07
Appendix
Table 10: DnR v2 test-set SI-SDR (dB) under matched inference-time boundary substitution. Avg. is the unweighted source mean.
Front end
SNR
ρ
RMS ratio
STFT (Mel-band control)
133.6
1.000
1.00
Learnable filterbank (K128)
-42.5
0.030
133.65
Appendix
Table 11: Identity reconstruction on the same mixed-signal probe with the separator bypassed. SNR is measured in dB; ρ is waveform correlation.
Figure 7: Peak-response frequencies of the 1025 paired analysis filters in the K128 learnable front end. Counts use a logarithmic vertical scale.
Figure 8: Input-dependent frequency partitions learned by DBS-BSRNN on DnR v2. Each row corresponds to one test mixture and shows the hard boundaries predicted for its central three-second window. Mixture order is shared by all panels. Boundary locations varied most strongly in B4, whereas B2 retained a comparatively stable layout.
Block
Unique per-block partitions
Cross-mixture replacement (%)
Adjacent-window replacement (%)
B1
64/64
27.56
22.13
B2
47/64
7.20
5.69
B3
64/64
26.41
19.62
B4
64/64
46.13
32.61
Appendix
Table 12: Learned boundary dynamics on DnR v2. Unique per-block partitions were counted among 64 central windows. Cross-mixture and adjacent-window columns report the mean percentage of internal boundaries replaced; mandatory block starts are excluded.
Figure 9: Empirical boundary-selection frequency over the central windows of 64 DnR v2 test mixtures. Mandatory block starts were excluded. B4 had the broadest input-dependent selection frequencies, whereas B2 was comparatively stable across mixtures.
Figure 10: Predicted boundaries across seven overlapping windows of one DnR v2 test mixture. Each row is a separate 3-second inference window and adjacent windows had a 1.5-second hop. The example had the median mean temporal replacement rate among the 64 analyzed mixtures. The partition was fixed within each window and changed between windows.
Figure 11: Mean input STFT power and predicted hard boundaries for four preselected DnR v2 examples in B3 and B4. Power was normalized to the maximum within each panel. The visualization supplies input context and does not by itself establish semantic alignment or a separation-quality gain.