Blind Source Separation Examples

Example files related to

are presented. Representative files are chosen from the database to allow for subjective separation quality evaluation.

Contents

The meeting room samples (medium reverberation) are gathered by the author, and originate from the public domain library librivox.
Data from the office is provided by B. Lösch [link].
The separated signal is decomposed according to [1]:
(1)
The scores are obtained from (1) as:
(3)
(4)

Notations

E1 = Separated target signal 1, see (1).
T1 = Original target signal 1.
I1 = Interference component, present in estimated target 1 signal (e_interf).
A1 = Artifact component, present in estimated target 1 signal (e_artif).
Likelihood = Output of the particle filter algorithm (Particle filtering distribution mean and standard deviation of particle set).

Track = The annotated track and the estimated track. Error metrics are also plotted.
Mix = A mixture signal from a single microphone channel.

TASK 1 – Two simultaneous stationary sources

Meeting room:

Likelihood, Track, Oracle, Mix

E1 T1 I1 A1

E2 T2 I2 A2

SDR

+8.83

+9.54

SIR

+16.00

+18.96

Office:

Likelihood, Track, Oracle, Mix

E2 T2 I2 A2

E6 T6 I6 A6

SDR

+9.21

+9.83

SIR

+15.77

+17.82

TASK 2 – Two simultaneous sources, one moving

Meeting room:

Likelihood, Track, Oracle, Mix

E1 T1 I1 A1

E7 T7 I7 A7

SDR

+9.18

+8.13

SIR

+18.38

+16.04

TASK 3 – Three simultaneous sources

Office:

Likelihood, Track, Oracle, Mix

E2 T2 I2 A2

E4 T4 I4 A4

E6 T6 I6 A6

SDR

+6.74

+5.79

+7.91

SIR

+12.06

+12.56

+13.66

Meeting room:

Likelihood, Track, Oracle, Mix

E1 T1 I1 A1

E2 T2 I2 A2

E4 T4 I4 A4

SDR

+7.29

+7.49

+7.46

SIR

+14.86

+14.73

+14.43


TASK 4 – Three simultaneous sources, one moving

Meeting room:

Likelihood, Track, Oracle, Mix

E1 T1 I1 A1

E4 T4 I4 A4

E5 T5 I5 A5

SDR

+6.31

+7.29

+5.62

SIR

+12.10

+14.46

+14.39


TASK 5 – Four simultaneous sources

Office:

Likelihood, Track, Oracle, Mix

E1 T1 I1 A1

E3 T3 I3 A3

E4 T4 I4 A4

E6 T6 I6 A6

SDR

+4.57

+2.56

+3.12

+6.92

SIR

+9.94

+7.46

+6.98

+13.05

Bibliography

[1] Vincent et al., Performance measurement in blind audio source separation, 2006 IEEE Transactions on Audio, Speech, and Language Processing, 14, pages 1162-1469.

Contact

Pasi Pertilä
Tampere University of Technology
Department of Signal Processing
P.O.Box 553
pasi pertila (at) tut fi

-21.12.2011 Page created
- 9.1.2012 Citation [1] added. Anechoic data is pending publishing request.
- 18.6.2012 Reviewer requested changes performed, samples updated.
- 3.9.2012 Publication URL updated.