How to remove the noise introduced by UVR Denoise?

[gritty ruin]

After I put my isolated vocals through UVR Denoise, it introduces it's own noise at the bottom of the spectrum. As you can see in the screenshot, at the bottom of the audio there is a thin line of noise that was not there before.

No idea how to remove that.

[high knoll]

No idea as I used spectral denoise more often instead of UVR's.
but the bottom line thingy seems that some DC offset-related thing screwed it, probably from doing normalize with "DC offset removal" in Audacity, which shouldn't be done.

[high knoll]

have once tried UVR denoise and didn't cause that issue tho

[autumn escarp]

I KNEW IT IT REALLY INTRODUCES HIS OWN NOISE

[autumn escarp]

I once input a noiseless audiofile song accidentally and it still showed me it had noise.It's very deceiving

[gritty ruin]

Yeah I feel the same. what I did is to add a high pass filter in Audacity and set it to 20hz. It removes the noise introduced by UVR Denoise

[high knoll]

you guys show me the UVR settings like this

[gritty ruin]

I'll redo denoise after enabling TTA. Had that disabled. Will update then 🙂

btw. UVR Denoise seems to cut off at 17khz. Before UVR Denoise my dataset was at around 20khz, now its at 17. Is that a problem?

[high knoll]

no need to DM me, just show your settings, since I can't reproduce your issue with my settings

or just use my settings to see if it works

[high knoll]

my settings also wont do cut off

[autumn escarp]

I should try that

[gritty ruin]

Okay so using "High End Process" will prevent UVR Denoise from cutting off at 17khz.

But every possible settings, even your settings, will lead to noise at 20hz.

[paper stratus]

Ayo? @gritty ruin level 2 !!!

[gritty ruin]

This is the raw file before UVR Denoise. There is little to no noise from 0 to 20hz.

This is after UVR Denoise, using the settings you posted. As you can see, no cutoff at 17khz, which is good.
However, noise at 0-20hz was introduced.

[high knoll]

what are the input & output audio format and their bitrate?

[gritty ruin]

Ohh you solved it. It actually was the bitrate that is set on the output. I was putting out 16bit by accident, noise is gone with 24bit 🙂 Thank you so much

[high knoll]

better use 32-bit float on both input & output, for uncompromised fidelity

[gritty ruin]

the step before uvr denoise is BS Roformer on MVSEP and that only return 24bit to me sadly.

[high knoll]

set wav output and it will be 32-bit float

[high knoll]

sorry for necroposting lol, but I have another thought:
before separating or starting processing the source audio, make sure it has zero DC offset. izotope RX waveform statistics can show the metric.

to remove the DC offset, in Audacity: Effect > Volume & Compression > Normalize, then check only "remove DC offset"

in Adobe Audition: Effects > Volume & Compression > Normalize (process), then check only "DC bias adjust: 0%"

otherwise too much DC offset from the source could possibly cause issues in further processings, including separation, like that "bottom line of noise"

[oak nest]

u can else use the eq module and eq said frequency out

[high knoll]

yeah it also works, iirc codename had ever said that

[oak nest]

it's common audio engineering stuff so yeah

[oak nest]

but eqing stuff out may mess with the wave forms

and have some clipping

before

after

[high knoll]

here's some comparison between directly separating source with DC offset (doesn't necessarily have the "bottom line" tho):

and separating source with DC offset removed before:

[oak nest]

if u use it on a song yes

cuz some sounds do go rather low

[high knoll]

also take a look on the instrumental stem as well (respectively):

[oak nest]

yeah that'll work better

but it still messes it up

[oak nest]

u dont really need to touch the high eneds

[high knoll]

also thanks to the early DC offset removal, it can somewhat make separation do better

[oak nest]

yeah it'll probably not confuse the other models but bs roformer handels it all nicely

[oak nest]

dunno if u can avoid the possible clipped samples thru just using the dc offset thing on audacity

ill need to try

[high knoll]

converting to 32-bit float and/or normalize below 0 dB, so not have to worry on clipping

[oak nest]

still says it clipped

a lot less though

[high knoll]

yeah clipping warning before converting back to PCM format

[oak nest]

Well it's in general bad, but the models know how to remove em so yeah

[wheat mesa]

@high knoll Hey do you know best way to remove DC offset in Izotope RX?

[mortal arrow]

You can remove DC offset in iZotope with "De-hum" module. (also any high-pass would do the job, too)
Also, this thread is almost 2 years old 😭