Audio Source Separation Using Variational Autoencoders and Weak Class Supervision

Source Code

Source code is available at GitHub.

Examples of the Separation Results

Mixture

3

Soft Mask 1

Original Source 1

Separated Source 1

SDR	10.00
SIR	12.97
SAR	13.25

9

Soft Mask 2

Original Source 2

Separated Source 2

SDR	9.37
SIR	11.24
SAR	14.26

Mixture

4

Soft Mask 1

Original Source 1

Separated Source 1

SDR	10.54
SIR	14.83
SAR	12.70

9

Soft Mask 2

Original Source 2

Separated Source 2

SDR	9.14
SIR	11.08
SAR	13.90

Mixture

2

Soft Mask 1

Original Source 1

Separated Source 1

SDR	13.65
SIR	15.80
SAR	17.84

6

Soft Mask 2

Original Source 2

Separated Source 2

SDR	14.94
SIR	18.17
SAR	17.82

Mixture

1

Soft Mask 1

Original Source 1

Separated Source 1

SDR	14.98
SIR	20.44
SAR	16.47

5

Soft Mask 2

Original Source 2

Separated Source 2

SDR	13.51
SIR	16.13
SAR	17.05

Mixture

3

Soft Mask 1

Original Source 1

Separated Source 1

SDR	18.08
SIR	20.66
SAR	21.59

5

Soft Mask 2

Original Source 2

Separated Source 2

SDR	19.45
SIR	23.44
SAR	21.68

Mixture

0

Soft Mask 1

Original Source 1

Separated Source 1

SDR	4.97
SIR	5.87
SAR	13.22

5

Soft Mask 2

Original Source 2

Separated Source 2

SDR	7.46
SIR	12.84
SAR	9.17

Mixture

8

Soft Mask 1

Original Source 1

Separated Source 1

SDR	8.87
SIR	9.98
SAR	15.74

9

Soft Mask 2

Original Source 2

Separated Source 2

SDR	12.68
SIR	17.09
SAR	14.72

Mixture

1

Soft Mask 1

Original Source 1

Separated Source 1

SDR	13.69
SIR	36.13
SAR	13.71

6

Soft Mask 2

Original Source 2

Separated Source 2

SDR	4.60
SIR	4.75
SAR	20.55

Mixture

0

Soft Mask 1

Original Source 1

Separated Source 1

SDR	12.49
SIR	14.14
SAR	17.68

1

Soft Mask 2

Original Source 2

Separated Source 2

SDR	15.49
SIR	20.63
SAR	17.11

Mixture

2

Soft Mask 1

Original Source 1

Separated Source 1

SDR	2.63
SIR	3.10
SAR	14.25

8

Soft Mask 2

Original Source 2

Separated Source 2

SDR	8.93
SIR	10.80
SAR	13.84