MUSDB-HQ Benchmark Dataset

• MUSDB-HQ is a high-fidelity, multitrack music separation dataset that provides uncompressed stereo tracks and isolated stems for vocals, drums, bass, and other sources.
• It sets a standard testing ground by using consistent train/val/test splits and measurable metrics like SDR and nSDR for evaluating diverse separation models.
• Its detailed structure and high-quality composition drive innovations in both hybrid and lightweight audio processing architectures in music source separation research.

The MUSDB-HQ benchmark dataset is a high-fidelity, multitrack music separation corpus widely used in the evaluation and development of source separation models. Created as an uncompressed, high-resolution counterpart to the original MUSDB18 dataset, MUSDB-HQ provides a standard reference for supervised music source separation research by offering professionally produced stereo tracks with isolated stems for key musical sources.

1. Dataset Structure and Composition

MUSDB-HQ consists of 150 stereo tracks, each provided at full bandwidth and in uncompressed format. For each track, the dataset contains four time-aligned stems representing the canonical sources of mixdown music: vocals, drums, bass, and other. These stems are derived from the original multitrack sessions and serve as ground truth references for supervised separation tasks. The dataset is subdivided into three sets: 86 tracks for training, 14 for validation, and 50 for evaluation purposes, a partitioning that is strictly adhered to in benchmarking protocols (Défossez, 2021).

The tracks span diverse popular music genres and production styles, reflecting the complexity and variability encountered in real-world audio separation scenarios.

2. Role in Music Source Separation Benchmarking

MUSDB-HQ functions as the primary benchmark for assessing music source separation models in computational audio research. Progress in model architectures—including time-domain, frequency-domain, and hybrid systems—has been catalyzed by its availability and the rigor of its reference splits.

Objective separation quality on MUSDB-HQ is typically measured using chunk-level Signal-to-Distortion Ratio (SDR), with the SDR metric sometimes computed in different variants. In recent competitions such as the Music Demixing Challenge, both traditional SDR and normalized SDR (“nSDR”) have been reported. The formula for SDR is:

$$\mathrm{SDR} = 10 \cdot \log_{10} \left( \frac{\sum_n \| s(n) \|^2 + \varepsilon}{\sum_n \| s(n) - \hat{s}(n) \|^2 + \varepsilon} \right)$$

where $s(n)$ denotes the ground-truth signal, $\hat{s}(n)$ is the separated estimate, and $\varepsilon$ is a stabilizing constant (Défossez, 2021).

MUSDB-HQ’s importance is emphasized by its centrality in evaluations for models such as Hybrid Demucs (Défossez, 2021), Moises-Light (Yun-Ning et al., 8 Oct 2025), and resource-efficient U-Net variants. Consistent splits and full-bandwidth data ensure direct comparability across publications and challenge submissions.

3. Impact on Model Development and Architectural Innovation

The uncompressed and high-quality nature of MUSDB-HQ has enabled the development and validation of advanced separation systems. For instance, the hybrid Demucs architecture—featuring dual-branch processing in both time and frequency domains—was specifically developed, tuned, and evaluated using MUSDB-HQ.

Performance improvements achieved and validated on MUSDB-HQ include:

• Hybrid Demucs realizing roughly 1.4 dB mean SDR improvement over its waveform-only predecessor across all sources (Défossez, 2021).
• Subjective preferences (overall quality rated at 2.83/5 for hybrid Demucs vs. 2.36/5 for waveform-only) directly attributed to MUSDB-HQ-based evaluations.
• Lightweight models such as Moises-Light demonstrate that by leveraging careful band-split design and parameter-efficient transformations, competitive separation performance (average SDR ≈ 9.96 dB) is attainable with an order of magnitude fewer parameters than previous systems (Yun-Ning et al., 8 Oct 2025).

MUSDB-HQ provides not only input data fidelity required for modern architectures (such as those using compressed residual branches, local attention, and singular value regularization (Défossez, 2021)), but also serves as the principal source for generating augmented or transformed datasets (e.g., musdb-XL-train, which applies commercial limiters to MUSDB-HQ stems for training de-limiter networks) (Jeon et al., 2023).

4. Methodological Considerations and Evaluation Protocols

Rigorous evaluation using MUSDB-HQ includes:

• Train/val/test splits faithfully followed across studies and competitions.
• Objective performance reported using SDR (often chunk-based, median aggregated).
• Subjective evaluations supplementing objective metrics, requiring expert listeners to rate quality and leakage across sources.

Data handling must account for the stereo format and sample-level alignment between mixture and source stems. For hybrid or multi-domain models (e.g., those combining spectrogram and waveform processing), careful dimension matching, padding, and inversion (e.g., ISTFT) are necessary to synchronize estimated and reference signals (Défossez, 2021).

Common protocols also include data augmentation, multi-resolution loss functions, and input windowing (e.g., 9-second segments for Moises-Light) to align with real-world audio production timelines and diverse temporal contexts (Yun-Ning et al., 8 Oct 2025).

5. Derived Datasets and Extensions

MUSDB-HQ serves as the foundation for derivative datasets aimed at auxiliary tasks. Notably, musdb-XL-train is created by randomly mixing or using original MUSDB-HQ stems, then processing the resultant mixtures with a commercial limiter plug-in. This augmented corpus is tailored for the training of music de-limiter networks, enabling new research directions such as sample-wise gain inversion and dynamic range restoration (Jeon et al., 2023).

Alternative datasets, such as MoisesDB, are sometimes combined with MUSDB-HQ to enhance training diversity and boost generalization, with downstream models tested on MUSDB-HQ as the standard reference for objective separation quality (Yun-Ning et al., 8 Oct 2025).

6. Limitations and Challenges

Several challenges and ongoing limitations are associated with the use of MUSDB-HQ:

• Strict domain alignment is required for hybrid-domain models to avoid artifacts at the mixture-stem interface (Défossez, 2021).
• The relatively limited corpus size compared to large modern datasets can limit the capacity for data-hungry architectures to fully realize their advantage, though this is partially mitigated by careful splitting, augmentation, and the use of companion datasets (Yun-Ning et al., 8 Oct 2025).
• Lack of standardized subjective assessment protocols introduces potential evaluation inconsistency, though expert human evaluations remain an integral component of confidence in separation quality (Défossez, 2021).
• The mixing and mastering conventions embodied in the corpus may not reflect recent trends dominated by heavy limiting or non-standard source definitions; extensions such as musdb-XL-train aim to address this specific gap (Jeon et al., 2023).

7. Significance and Continuing Influence

The MUSDB-HQ dataset is a cornerstone of music source separation research, providing a consistent, high-fidelity, and openly available reference for empirical evaluation. Its structure and reference protocols drive advances in hybrid architectures, lightweight design, dynamic range restoration, and evaluation standards. Innovations validated on MUSDB-HQ have demonstrated measurable improvements in both objective (e.g., SDR, nSDR) and subjective (expert rating) criteria. Its role as a benchmarking authority continues to shape state-of-the-art methodology and the direction of computational audio research (Défossez, 2021, Jeon et al., 2023, Yun-Ning et al., 8 Oct 2025).