MUSDB18 - a corpus for music separation

By Z. Rafii, Antoine Liutkus, Fabian-Robert Stöter, S. I. Mimilakis, Rachel M. Bittner. 2017

437 citations recorded by the source corpus.

Subfields: Speech & Audio, Signal Processing

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Speech & Audio

In Signal Processing

Citations in the corpus

High-Fidelity Audio Compression with Improved RVQGAN (2023, methodology): “For evaluation, we use the evaluation segments from AudioSet [14], two speakers that are held out from DAPS [26] (F10, M10) for speech, and the test split of MUSDB [31].”
Vocos: Closing the gap between time-domain and Fourier-based neural vocoders for high-quality audio synthesis (2023, methodology): “In this context, we evaluate the performance of Vocos with out-of-distribution audio using the MUSDB18 dataset (Rafii et al., 2017), which includes a variety of multi-track music audio like vocals, drums, bass, and other instruments, along with the original mixture.”
Differentiable Model Compression via Pseudo Quantization Noise (2021, methodology): “We train a Demucs source separation model (Défossez et al. 2019) (MIT license) with a depth of 6 and 64 initial hidden channels, on the MusDB dataset (Rafii et al. 2017)2, which is released under mixed licensing3.”
Music Source Separation in the Waveform Domain (2019, positive): “We present experiments on the MusDB [Raﬁi et al., 2017] benchmark.”
Wave-U-Net: A Multi-Scale Neural Network for End-to-End Audio Source Separation (2018, methodology): “75 tracks from the training partition of the MUSDB [17] multi-track database are randomly assigned to our training set, and the remaining 25 tracks form the validation set, which is used for early stopping.”
Singing Voice Separation: A Study on Training Data (2019, methodology): “Indeed, public training datasets have been regularly released (MIR1K [17], MedleyDB [18], DSD100 [19] MUSDB [20]) and used to compare different methods, but they are rather small, and often lack diversity.”
Parallel and Flexible Sampling from Autoregressive Models via Langevin Dynamics (2021, result): “We do not compare results on the popular MusDB dataset (Raﬁi et al., 2017) because this dataset has insufﬁcient single-channel audio to train WaveNet generative models.”
Joint Phoneme Alignment and Text-Informed Speech Separation on Highly Corrupted Speech (2020, methodology): “We use the instrumental accompaniments of the MUSDB18 data set [22] as music signals and mix them with the TIMIT [23] speech signals.”
Improving singing voice separation using Deep U-Net and Wave-U-Net with data augmentation (2019, background): “For all experiments, the starting point is the musdb18 dataset [24].”
MERT: Acoustic Music Understanding Model with Large-Scale Self-supervised Training (2023, methodology): “MUSDB18 contains 150 full-length music tracks (˜10 hours), along with multiple isolated stems.”
SingSong: Generating musical accompaniments from singing (2023, positive): “To evaluate model performance on isolated vocals, we adopt the MUSDB18 dataset (Rafii et al., 2017), which contains 10 hours of professional, studioisolated vocal and instrumental “stems” for 150 source mixes.”
Hybrid Transformers for Music Source Separation (2022, methodology): “The reference dataset that is used to benchmark MSS is MUSDB18 [3, 5] which is made of 150 songs in two versions (HQ and non-HQ).”
Meta-Learning Extractors for Music Source Separation (2020, methodology): “[5] Zafar Rafii, Antoine Liutkus, Fabian-Robert Stöter, Stylianos Ioannis Mimilakis, and Rachel Bittner, “The MUSDB18 corpus for music separation,” Dec. 2017.”
The 2018 Signal Separation Evaluation Campaign (2018, methodology): “For the organization of the present SiSEC, the MUSDB18 corpus was released [21], comprising tracks from MedleyDB [4], DSD100 [19, 14], and other material.”
Wavesplit: End-to-End Speech Separation by Speaker Clustering (2020, neutral): “This situation does not occur when separating different musical instruments from predefined categories [6], or separating speech from non-speech noise [7].”
Spleeter: a fast and efficient music source separation tool with pre-trained models (2020, methodology): “The performance of the pre-trained models are very close to the published state-of-the-art and is one of the best performing 4 stems separation model on the common musdb18 benchmark (Rafii, Liutkus, Stöter, Mimilakis, & Bittner, 2017) to be publicly released.”
Upsampling Artifacts in Neural Audio Synthesis (2020, methodology): “[28] Zafar Rafii, Antoine Liutkus, Fabian-Robert Stöter, Stylianos Ioannis Mimilakis, and Rachel Bittner, “The MUSDB18 corpus for music separation,” 2017.”