Music source separation is the task of isolating the instrumental tracks from a music song. Despite its spectacular recent progress, the trend towards more complex architectures and training protocols exacerbates reproducibility issues. The band-split recurrent neural networks (BSRNN) model is promising in this regard, since it yields close to state-of-the-art results on public datasets, and requires reasonable resources for training. Unfortunately, it is not straightforward to reproduce since its full code is not available. In this paper, we attempt to replicate BSRNN as closely as possible to the original paper through extensive experiments, which allows us to conduct a critical reflection on this reproducibility issue. Our contributions are three-fold. First, this study yields several insights on the model design and training pipeline, which sheds light on potential future improvements. In particular, since we were unsuccessful in reproducing the original results, we explore additional variants that ultimately yield an optimized BSRNN model, whose performance largely improves that of the original. Second, we discuss reproducibility issues from both methodological and practical perspectives. We notably underline how substantial time and energy costs could have been saved upon availability of the full pipeline. Third, our code and pre-trained models are released publicly to foster reproducible research. We hope that this study will contribute to spread awareness on the importance of reproducible research in the music separation community, and help promoting more transparent and sustainable practices.
Figures & tables
Vocals
Bass
Drums
Other
Average
Training on MUSDB18-HQ
CWS-PResUNet [ 29 ]
8.92
5.93
6.38
5.84
6.77
KUIELab-MDX-Net [ 30 ]
8.97
7.83
7.20
5.90
7.47
Hybrid Demucs [ 12 ]
8.35
8.43
8.12
5.65
7.64
HT Demucs [ 15 ]
7.93
8.48
7.94
5.72
7.52
TABLE I: Separation performance (chunk SDR in dB) for state-of-the-art MSS models on the MUSDB18-HQ test set.
Fig. 1: Overview of BSRNN and its variants (A). Blocs with dashed line contours denote variants not in the original architecture. The residual network (B) in the sequence / band modules are based on either RNNs, as in the original model, or dilated convolutions.
Validation uSDR (dB)
# Parameters (M)
Energy (kWh)
Vocals
Bass
Drums
Other
Average
codecarbon
green algo.
1
Base model: N=64 , R=8
7.7
6.1
9.7
4.8
7.1
32.3
127
168
Training parameters
2
Accumulating gradients
8.0
5.8
9.6
4.9
7.1
-
129
170
3
Monitoring with the loss
7.5
6.4
9.3
4.8
7.1
-
120
159
4
Loss domain: time
7.9
6.1
9.4
4.9
7.1
-
116
153
TABLE II: Comparison of various models’ performance: best model’s validation uSDR (in dB), total number of parameters (in millions, “-” denotes the same value as in the preceding line), and estimated energy for training all sources (in kWh).
Fig. 2: Validation uSDR over epochs for the base model on the vocals track, with a patience of 10 (left) or 30 (right). Each color corresponds to a different run, and the dashed lines correspond to each run’s best uSDR.
Fig. 3: Performance vs. training energy (estimated using the Green Algorithms calculator) for model variants, where each number marker corresponds to the line number in Table II . Blue dots highlight models that are optimal in a Pareto sense.
Vocals
Bass
Drums
Other
Average
uSDR
cSDR
uSDR
cSDR
uSDR
cSDR
uSDR
cSDR
uSDR
cSDR
Paper’s results [ 10 ]
10.04
10.01
6.80
7.22
8.92
9.01
6.01
6.70
7.94
8.24
Our implementation
Fader
9.16
9.14
6.51
7.72
8.55
8.07
5.44
5.68
7.42
7.65
OLA, hop = 1.5 s
9.10
-
6.47
-
8.33
-
5.44
-
7.34
-
OLA, hop = 0.5 s
9.18
-
6.60
-
8.43
-
5.53
-
7.44
-
TABLE III: Source separation performance on the MUSDB18-HQ test set (uSDR and cSDR, in dB).
Band-split recurrent neural network (BSRNN) is a popular music source separation model that yields close to state-of-the-art results using reasonable computational resources and public datasets. It is therefore interesting from a reproducible research perspective, but achieving its performance is not straightforward since its full code is not available. In this paper, we conduct a replication of BSRNN via implementing the full pipeline. We extend the original paper's analysis by experimentally studying various design choices about data preprocessing, the optimization protocol, and architectural parameters. We report and discuss this project's energy cost, and we underline how its footprint could have been substantial lower upon availability of the full pipeline, which advocates for more reproducible research practices. To comply with this objective, we publicly release our code and pre-trained models.
Paul Magron, Romain Serizel, Constance Douwes
Université de Lorraine, CNRS Inria, LORIA, F-54000 Nancy, France · Centrale Med, Aix Marseille Univ CNRS, LIS, Marseille, France
The Inaugural Music Source Restoration (MSR) Challenge targets the recovery of original, unprocessed stems from fully mixed and mastered music. Unlike conventional music source separation, MSR requires reversing complex production processes such as equalization, compression, reverberation, and other real-world degradations. To address MSR, we propose a two-stage system. First, an ensemble of pre-trained separation models produces preliminary source estimates. Then a set of pre-trained BSRNN-based restoration models performs targeted reconstruction to refine these estimates. On the official MSR benchmark, our system surpasses the baselines on all metrics, ranking second among all submissions. The code is available at https://github.com/xinghour/Music-source-restoration-CUPAudioGroup
Xinlong Deng, Yu Xia, Jie Jiang
College of Artificial Intelligence, China University of Petroleum, Beijing
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
Bingzhe Zhang, Kai Li, Xiaolin Hu
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