Musical Source Separation: An Introduction

V. EVALUATION OF MUSICAL SOURCE SEPARATION MODELS

 

Data Sets

  • DSD100: The Demixing Secrets Dataset 100 (DSD100) consists of a total of 100 full-track songs of different styles and includes the mixtures and four original sources/stems.
  • MedleyDB: The MedleyDB is a dataset of 46 annotated, royalty-free multitrack recordings.
  • MUSDB18: A description of the data set and the audio files used for the 2018 SiSEC Evaluation campaign.
  • Others: A collection of data sets used for MSS.
 

Listening Tests Resources

  • MUSHRAM: A Matlab interface for MUSHRA listening tests in Matlab.
  • webMUSHRA: Web-based MUSHRA compliant to the ITU-R Recommendation BS.1534 (MUSHRA).
 

Evaluation Metrics Resources

  • MUSEVAL: Evaluation Toolkit in Python for bss_eval v4.
  • mir_eval.separation: Python implementation of bss_eval v3.
  • PEASS: Matlab implementation of PEASS Toolkit.
 

SiSEC Evaluation Campaign

  • MUS18: The community-based signal separation evaluation campaign offers a task for professionally-produced music recordings.
  • Results SiSEC 2018: The full description of the results of the SiSEC 2018 Evaluation Campaign