Investigating the Performance and Energy Costs of Replicating Band-Split RNN for Music Source Separation

Paul Magron; Romain Serizel; Constance Douwes

cs.SDSep 18, 2026

Investigating the Performance and Energy Costs of Replicating Band-Split RNN for Music Source Separation

Authors: Paul Magron, Romain Serizel, Constance Douwes

Organizations: Université de Lorraine, CNRS Inria, LORIA, F-54000 Nancy, France · Centrale Med, Aix Marseille Univ CNRS, LIS, Marseille, France

Abstract

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.

Figures & tables

	Vocals	Bass	Drums	Other	Average
KUIELab-MDX-Net [ 8 ]	8.97	7.83	7.20	5.90	7.47
Hybrid Demucs [ 3 ]	8.35	8.43	8.12	5.65	7.64
HT Demucs [ 4 ]	7.93	8.48	7.94	5.72	7.52
TFC-TDF UNet v3 [ 9 ]	9.59	8.45	8.44	6.86	8.34
BSRNN [ 10 ]	10.01	7.22	9.01	6.70	8.24
SIMO-BSRNN [ 11 ]	9.73	7.80	10.06	6.56	8.54

Table 1: Separation performance (cSDR in dB, see Section 3.3 ) for state-of-the-art MSS models trained and tested on MUSDB18-HQ.

		Validation uSDR					Parameters	Energy
		Vocals	Bass	Drums	Other	Average		codecarbon	GA
1	Base model: $N=64$ , $R=8$	$7.7$	$6.1$	$9.7$	$4.8$	$7.1\phantom{\star}$	32.3	127	168
2	Loss domain: time	$7.9$	$6.1$	$9.4$	$4.9$	$7.1\phantom{\star}$	-	116	153
3	Loss domain: STFT	$7.9$	$6.4$	$9.6$	$4.9$	$7.2\phantom{\star}$	-	131	173
4	Accumulating gradients	$8.0$	$5.8$	$9.6$	$4.9$	$7.1\phantom{\star}$	-	129	170
5	Monitoring with the loss	$7.5$	$6.4$	$9.3$	$4.8$	$7.1\phantom{\star}$	-	120	159

Table 2: Model comparison in terms of best validation uSDR (in dB), number of parameters (in millions, “-” denotes the same value as in the preceding line), and estimated energy for training all sources (in kWh). Each line describes the difference with the base model at line 2 . Large deviations from the base average uSDR, as discussed in Section 4.1.1 , are indicated with the ⋆ symbol.

Figure 1: Training the base model on the vocals track, with a patience of 10 (left) or 30 (right). Each color corresponds to a different run/seed, and the dashed lines correspond to each run’s best uSDR.

Figure 2: Performance vs. training energy for small-size model variants, where each marker corresponds to the line number in Table 2 .

Audio Source Separation

Audio source separation isolates individual speakers, instruments, or sound events from mixtures for listening, analysis, and audio editing.

See all 65 papers

The Costs of Reproducibility in Music Separation Research: a Replication of Band-Split RNN
Mar 10, 2026Paul Magron, Romain Serizel, Constance DouwesMusic Source Separation Audio Source Separation
Université de Lorraine, CNRS, Inria, LORIA, F-54000 Nancy, France · Centrale Med, Aix Marseille Univ, CNRS, LIS, Marseille, France
Learning Where to Split: Dynamic Band Splitting for Audio Source Separation
Oct 10, 2026Bingzhe Zhang, Kai Li, Xiaolin HuAudio Source Separation
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
Music Source Restoration with Ensemble Separation and Targeted Reconstruction
Mar 13, 2026Xinlong Deng, Yu Xia, Jie JiangAudio Source Separation Music Source Separation
College of Artificial Intelligence, China University of Petroleum, Beijing

Mar 10, 2026J/K move · Enter open · S save

The Costs of Reproducibility in Music Separation Research: a Replication of Band-Split RNN

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

Music Source Separation Audio Source Separation ML Reproducibility Recurrent Neural Networks

Open paper AskPDF arXiv

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.

Abstract

Audio Source Separation

Explore similar work

The Costs of Reproducibility in Music Separation Research: a Replication of Band-Split RNN

Learning Where to Split: Dynamic Band Splitting for Audio Source Separation

Music Source Restoration with Ensemble Separation and Targeted Reconstruction