Skip to content

Spectograms too bleedy #31

Description

@lucellent

Image

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

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions