/ccom_mdx2023

MIT LicenseMIT

ccom_mdx2023

This is the repository of team CCOM for Sound Demixing Challenge 2023. For Music Demixing Track, We get rank #1 in the Leaderboard A and rank #4 in the Leaderboard B.

Requirements

  • For inference and bleeding training, see apt.txt and requirement.txt in the labelnoise_submission branch.
  • For labelnoise training, you need to clone DEMUCS and install all its dependencies, with also packages in apt.txt and requirement.txt needed.

Methods

  • For Leaderboard A, we used loss truncation in pre-training phase and then we filtered the dataset automatically with the pretrained loss truncation model. For further details please see README.md in the labelnoise_train branch.
  • For Leaderboard B, we designed a multi-spectrogram U-Net model. Details are shown in the bleeding_train branch.

Results

Leaderboard A

  • Submission ID: 219610

  • Submitter: dramatic

  • Final rank: 1st place

  • Final scores:

    Mean SDR SDR Bass SDR Drums SDR Other SDR Vocals
    7.455 8.117 7.993 5.342 8.369

Leaderboard B

  • Submission ID: 217439

  • Submitter: liujiafeng

  • Final rank: 4th place

  • Final scores:

    Mean SDR SDR Bass SDR Drums SDR Other SDR Vocals
    6.203 6.337 6.323 4.284 7.867