Sound samples for

Notes on nonnegative tensor factorization of the spectrogram for audio source separation:
statistical insights and towards self-clustering of the spatial cues

Cédric Févotte (CNRS LTCI; Télécom ParisTech) and Alexey Ozerov (IRISA, Rennes)


   The original sound files used in these experiments were obtained from the Signal Separation Evaluation Campaign (SiSEC) website.

wdrums
mix
  (Hi-hat)
s1
(Drums)
s2
(Bass)
s3
  KL-NTF.mag
SDR -0.2 0.4 17.9
ISR 15.5 0.7 31.5
SIR 1.4 -0.9 18.9
SAR 7.4 -3.5 25.7
  KL-cNTF.mag
SDR -0.02 -14.2 1.9
ISR 15.3 2.8 2.1
SIR 1.5 -15.0 18.9
SAR 7.8 13.2 9.2
  IS-NTF.pow
SDR 12.7 1.2 17.4
ISR 17.3 1.7 36.6
SIR 21.1 14.3 18.0
SAR 15.2 2.7 27.3
  IS-cNTF.pow
SDR 13.1 1.8 18.0
ISR 17.0 2.5 35.4
SIR 22.0 13.7 18.7
SAR 15.9 3.4 26.5

nodrums
mix
  (Bass)
s1
(Lead G.)
s2
(Rhythmic G.)
s3
  KL-NTF.mag
SDR 13.2 -1.8 1.0
ISR 22.7 1.0 1.2
SIR 13.9 -9.3 6.1
SAR 24. 2 7.4 2.6
  KL-cNTF.mag
SDR 5.8 -9.9 3.1
ISR 8.0 0.7 6.3
SIR 13.5 -15.3 2.9
SAR 8.3 2.7 9.9
  IS-NTF.pow
SDR 5.0 -10.0 -0.2
ISR 7.2 1.9 4.2
SIR 12.3 -13.5 0.3
SAR 7.2 3.3 -0.1
  IS-cNTF.pow
SDR 3.9 -10.2 -1.9
ISR 6.2 3.3 4.6
SIR 10.6 -10.9 -3.7
SAR 3.7 1.0 1.5