when you make a new notebook in google colab it has a prompt feature at the bottom of the notebook.




So i prompted:

- I want to pull this github url: https://github.com/NVIDIA/diffusion-audio-restoration.git
- create a requirements.txt file based on the repo
- download these models and place them in the models folder:
- https://huggingface.co/nvidia/audio.../main/ckpt/A2SB_onesplit_0.0_1.0_release.ckpt
- https://huggingface.co/nvidia/audio.../main/ckpt/A2SB_twosplit_0.5_1.0_release.ckpt
- https://huggingface.co/nvidia/audio.../main/ckpt/A2SB_twosplit_0.0_0.5_release.ckpt
- create a gradio interface for inference with audio upload and a drop down choice for the models


The cool thing is it uses gemini to fix any errors and you can keep chatting with it to refine your build. It does a pretty good job knowing what to put in the requirements.txt I just think there are versioning issues with some py libs.

 

It doesn't show up on my layout. But Google's AI on Google search also errors out so maybe it's something to do with my strict network blocking / filtering. Thank you for the assistance anyways. You were able to run it on Colab when I couldn't. I always knew Google was racist and biased.

 

I dont think its racist? I know hella mixed races that use colab all the time.

Yeah i was able to run it on colab.

It might be because i pay for gemini that its included? I just check my girl has it too.

 

did you click on the toggle gemini button in the middle at the bottom?

 

As always I've of course posted it (and not only) on my [REQ] Delossifier function discussion @ OpenVINO™ AI Plugins for Audacity repo...

 

About some time ago, there was a guy here who transformed VLC's DSP into a VST plugin. I have Sony's DSEE-HX (premium version) that shipped with Sony hardware only. If anyone has the knowhow to create a VST plugin from it, please PM me.

The one I have was ripped from Sony Xperia XZ2 Sound System and had API functions that prevented it from being used on nonproprietary devices removed.

 

There's an interesting discussion (with intrumental tests/measures) about DSEE HX AI / Ultimate / Extreme Do @ ASR...

...maybe you can profile it using NAM training...

Anyway you can also try to "vibe code" it or:

Another option is to ask some specialized group (such as R2R, P2P, TCD, etc.) to mod/release it unlocked.

 

Last edited: Aug 18, 2025

Sorry, I don't know any coding, this is beyond my skills and I have no way to contact the groups. If I remember correctly, TCD use to share their email in their nfo's but it doesn't work, email always bounces back returning bad address. Thank you for the tip though, yes it will be very good if somebody could port it into a vst plugins but alas! somethings remains out of reach forever.

 

You dont need to know how to code to "vibe code" but that another thing.

if yall want to try DSEE you can download this free app from sony.

https://www.sony.com/electronics/support/articles/00257100




you can run apps to record the output of the algo.

Ill still stand by using Theimo Stereo Tool over a lot of these algos. Ive tested them all and if you dive in deep with Stereo tool you can get spectrograms like all listed in this thread + 50 other options to improve low quality BR music.

Its also on sister site for all platforms even command line.

 

It seems more like marketing than anything else "at first glance" IMHO: if the DSEE HX function had been valid "technology" I think they would have included it in SpectraLayers Pro and/or Sony Catalyst Production Suite too...

 

Yea, that's what I do for the stuff that can't ever be found in better quality and I want it badly enough. But I'm skeptical about your claim of Stereo Tool being better. My results indicate that DSEE is ages ahead of anything similar, Stereo tool or otherwise.

You mind sharing some Stereo tool upscaling presets so I can give the tests one more go ? If possible.

Thank you for the help and response. Your contribution is highly appreciated.

 

It is more than valid tech, it just doesn't write new frequencies like so many other tools that you can see in the spectrogram but it actually adds the information that you can tell by the difference how the processed music sounds. The only think that ever made a difference in hearing tests is actually DSEE, all the rest is just noise until proven otherwise barring Nvidia's tech, haven't tested that so can't say.

 

yep here are some that I like:
https://pixeldrain.com/u/znnwQZFG

but its all dependent on the source material so dive in and move some sliders.

 

Here is a screenshot of the before and after.

you should go into processing after and tweak the high settings to really hear the difference. The High Frequency tools post delossifer use the information from that to enhance those chopped up tops.

 

Thank you, I will most definately give it a try and post the results. Could you please tell me what version of Stereo Tool is in your screenshot ?

 

I couldn't change the highlighted drop down, with or without your presets. Not sure if it's bug or something wrong with the host. Anyways can't use this without being able to change that.

 

maybe download the recent beta:

https://forums.stereotool.com/viewtopic.php?p=78829&sid=d64adee58f26ae08576a126cf37fab4f#p78829

PM me if that doesnt work.

 

It would be very interesting to perform some null test (delossyfied vs original/lossy) to better understand the fidelity of various approaches...