SLP12 Possible bug in Unmix Multiple Voices - delayed layer start gets baked into export

@Robin_Lobel
I would like to report what seems to be a reproducible bug in SpectraLayers 12 Pro v12.0.30.

My workflow is based on interview recordings with 2 mono tracks:

  • Track A01: lavalier mic on the reporter/interviewer, mainly containing the reporter’s questions

  • Track A02: reportage/handheld mic, mainly containing the interviewee’s answers

Of course, on both tracks the other person’s voice is also faintly present as bleed.

For further processing, I do the following:

  1. Import the video footage into DaVinci Resolve

  2. Export both audio tracks as 32-bit float BWAV, preserving the video timecode

  3. Import both BWAV files into SpectraLayers 12 Pro, each into its own project

  4. Normalize to -1.5 dB

  5. Run Unmix Noisy Speech

  6. Check the Noise layer for wrongly separated speech remnants and delete those from that layer if necessary

  7. On the Speech layer, select a section where both voices are audible

  8. Open Unmix Multiple Voices

  9. First click Register Multiple Voices

  10. After the voices have been registered, apply Unmix Multiple Voices

  11. Delete the layer containing the unwanted voice (the bleed)

  12. Merge all remaining layers

  13. Export the mixdown

The problem:

A reproducible timing error occurs after using Unmix Multiple Voices.

If I place the cursor at the very start of the file (for example with the Home key), I can see that one of the resulting voice layers cannot be positioned at 00:00 anymore, because SpectraLayers seems to treat that layer as if its audio begins a few seconds later. This is also visible visually in the layer itself.

This always affects the layer containing the voice that starts later than the other speaker. The other voice layer behaves correctly.

As soon as I delete the unwanted layer and then merge the remaining layer, this offset gets effectively baked in and is also present in the exported file.

So the export is no longer correctly aligned to the original file start / time reference.

Expected result:
After Unmix Multiple Voices, all resulting layers should remain aligned to the original file start, and merging/exporting should preserve the correct timing.

Actual result:
The layer containing the later-starting voice appears shifted so that it no longer starts at 00:00, and after merge/export this timing offset is written into the exported file.

I can provide the affected audio files for reproduction if needed.

System specs:

  • SpectraLayers 12 Pro 12.0.30

  • Windows 10 Pro 22H2

  • AMD Ryzen 9 5950X (16 cores)

  • Gigabyte B550 Vision D-P

  • Corsair Vengeance LPX 128 GB (4 x 32 GB) DDR4 3600 MHz C18

  • PNY NVIDIA RTX A6000 (Ampere) 48 GB

  • NVIDIA WHQL Studio Driver 553.50

  • Seasonic PRIME TX-1000 80PLUS Titanium 1000 W

  • WD_BLACK SN850 2 TB NVMe PCIe Gen4 (system)

  • Crucial P5 2 TB (fast cache)

  • 4 x Crucial MX500 4 TB as 16 TB RAID 0 (big cache)

  • LSI 8-port RAID controller

  • 8 x 16 TB Toshiba Enterprise MG08ACA16TE as 104 TB RAID 5 (work files)

Thanks.

You can see that the yellow layer got a start cutoff/offset and you can’t go before that with your cursor:

1 Like

Anyone able to confirm this bug?

You might have some personal usage reason for staying in that version… However, have you tested this on your system, same configuration and latest SL 12.0.40 release?

I saw no reason to update as my hardware does not struggle with too less VRAM.

The release notes for 12.0.40 show no signs of a fix for this particular filter:

Improvements

  • GPU and RAM management has been optimized.

Miscellaneous

  • Pressing the ESC key now properly cancels a running process.

  • Moving a layer in timeline with the Transform tool no longer creates additional silence in some cases.

  • The preference “Display > Selected Layers” now works as expected.

  • Voice DeClip no longer generates low frequency audio artifacts.

1 Like

@Robin_Lobel
Can you confirm the bug? I’ve also sent you a PM with a link to test files.

@Robert_Niessner Yes, sorry for the late answer !

This is rather a bad design decision than a bug, I’ll make sure this get improved.
Meanwhile the workaround I suggest is to mute the layer you don’t want rather than deleting it. This will make sure the sum of all layers always starts at 0 on the timeline (taking into account the muted layer).

2 Likes