(Possible) Faulty segmentation on Mel-Roformer-Karaoke-Aufr33-Viperx model?

[frnksmdlkedjnfr]

Dear Dev.

Mel-Roformer-Karaoke-Aufr33-Viperx has always been my fav model for its outstanding seperation performace. But after I reinstalled my Windows and the whole UVR5-UI, I began to notice some weird separation results. To describe it, it sounds like the song had been cut into non-consecutive segments and processed, leaving weird gaps where the unprocessed audio remained.

I reckon this had something to do with audio segmentation. I fiddled with the segment size, override segment size, and overlap, but just couldn't get it right. As far as I've tested, this is the only model that has this problem.

I'll attach the original song and the separated instrumental track. In the instrumental track, at 1:19 you can clearly hear the lead vocal suddenly was taken out by the model, and just seconds later at 1:33, the same phrase was processed perfectly. I wonder if you'll be kind enough to test it on your local machine and see if the model behaved the same. This problem doesn't just apply to this one song.

[frnksmdlkedjnfr]

Update: Tried using MSST-WebUI to load the same model, got completely normal results. Did the new update break something?

[Eddycrack864]

Sorry for the late reply. I think you may have sent me the wrong sample, because the sample you sent me is from the "BS-Roformer-Viperx-1297" model and not the "Mel-Roformer-Karaoke-Aufr33-Viperx." In any case, I tried it and noticed that, just as you told me, at 1:13 the main vocals were taken, and then at 1:33 they weren't.

I assume this is the problem you're having, but I have a question regarding what you say here "To describe it, it sounds like the song had been cut into non-consecutive segments and processed, leaving weird gaps where the unprocessed audio remained", let me explain in detail:

I've noticed that the "Mel-Roformer-Karaoke-Aufr33-Viperx" model uses a different config file than the one used by UVR5 UI, but when I use the new config file (obtained from MSST) the results are similar to what you describe, so I'm not sure, that's why I'll send you both results so you can tell me what problem you're having exactly.

https://drive.google.com/drive/folders/1CHkOtGHYIXCmnrDR0v3QEjWYWA-zmXfs?usp=drive_link

In "Mel-Roformer-Karaoke-Aufr33-Viperx" folder there are the UVR5 UI separated one and the MSST separated one and in "unsure" folder there are the separated lead/backvocals with MSST config.