Is there anything that can be done either architecturally or changing the parameters of the model, to improve how it handles bleed? This image is coming from a model trained for both vocals and instrumental, but the instrumental clearly has a lot of vocal residues that are stubborn:
band_stride:
- 1
- 4
- 16
band_kernel:
- 3
- 4
- 16
conv_depths:
- 3
- 2
- 1
Changing these to the following values barely has an effect except for higher SDR scores:
band_stride:
- 1
- 4
- 8
band_kernel:
- 3
- 4
- 8
conv_depths:
- 3
- 3
- 3
You can also see a horizontal line on the upper frequencies, where the boundary of the second and third bands are, sometimes messing up transients
Is there anything that can be done either architecturally or changing the parameters of the model, to improve how it handles bleed? This image is coming from a model trained for both vocals and instrumental, but the instrumental clearly has a lot of vocal residues that are stubborn:
band_stride:
- 1
- 4
- 16
band_kernel:
- 3
- 4
- 16
conv_depths:
- 3
- 2
- 1
Changing these to the following values barely has an effect except for higher SDR scores:
band_stride:
- 1
- 4
- 8
band_kernel:
- 3
- 4
- 8
conv_depths:
- 3
- 3
- 3
You can also see a horizontal line on the upper frequencies, where the boundary of the second and third bands are, sometimes messing up transients