AI Delay Removal / Dry Signal Extraction

[exit-eternity]

Scenario: single-channel instrument (say a piano) with heavy delays.
Problem: if you try to get MIDI out of it via Melodyne or whatever, the delays make a total mess. Current AI tools (RX, SpectraLayers) are great at stem separation, but they don’t touch repeated echoes/delay tails.

So here’s the thought:

Identify the dry hits (main notes).

Detect repeated echoes/delays.

Suppress/remove them while keeping the dry signal intact.

Even a janky version would be awesome. You could even give hints: BPM, delay length (1/8, 1/16), ping-pong/mono… basically just make it TRY.

Has anyone played with this? Any plugin, script, or workflow hacks that approximate this? Even if it’s messy, it’d still be massively useful for cleaning up delayed instruments for MIDI or processing.

Anyone?

[Aloysius]

If I was trying to get the midi from it, I think I'd use a noise gate before melodyne.

[VariKusBrainZ]

You could try a mid/side tool in the hope theres more uneffected on the mid channel, just a thought

[exit-eternity]

edit:
lets assume killing side channel helps; but not fully! theres still tons of delays left.

lets assume a noise gate doesnt work, since the delays wont be so much lower in amplitude that setting the threshold of the gate would make a decent result.
Yes, both of these suggestions are ofc great, but lets assume ive done this already, but what im after is a bit of a cleaner result!

[Aloysius]

In the old days we used to work things out by listening and trying to play what we heard. You could try that.

[Thymos]

Maybe slowing down the bpm and cutting echoes out of the sample manually to increase the precision of the midi detection afterwards?

[Thymos]

Crazy idea: You could create a copy of the sample on a separate channel and align the first transient to the second transient of the original sample. Then you flip the phase of the second sample by 180 degree and use phase cancelation to remove the echoes.

[swilow11]

In the old days we used to work things out by listening and trying to play what we heard. You could try that.

Awesome but also not really relevant here.

[Aloysius]

Did you like my other suggestion?
Do you have a suggestion you would like to share yourself or are just policing the thread?

[Tj Shredder]

I think its possible wit AI. Just look around which developer has experience in training their own tool. They might think its worth to do it... Out of the answers I guessed it doesn't exist yet, but I just asked the AI and it pointed me to the following existing tools, online echo removal:
https://www.lalal.ai/echo-reverb-remover/
https://voice.ai/tools/echo-remover
Let us know if they are useful if fed with an echoed stem...
If its not at all useful, fire a feature request to Steinberg and RipX... I bet they want to be better than the competition...

[osiris]

I don't think it can be done completely. I have a plugin called Deverb and it does a good job but you still get ghosts. I see a lot of people talking about the Acon Digital plug.

[lingyai]

Did you like my other suggestion?
Do you have a suggestion you would like to share yourself or are just policing the thread?

His comment is no more "policing" than yours

[lingyai]

In the old days we used to work things out by listening and trying to play what we heard. You could try that.

In the old days we didn't use computers either. Seems you're on the wrong site.

[Ah_Dziz]

It's probably a pretty tricky problem for any signal separation system. If it was a perfectly clean digital delay with noticeably higher level on the dry signal you could have a good shot at it. It's already extra tricky to split signals that are at the same pitch with wildly different timbre (like spitting a vocal ensemble into individual voices). Approaching it based off of the delay time seems the best way provided there is no large modulation of the delayed material. Unfortunately I only really use these kinds of things for stripping apart samples. I don't know which of any of them allow you to really muck about with the underlying detection and extraction models. I'd think cleaning it up to the point that you could easily hear the original clearly should be doable but I'm not sure about having it clean enough to pull polyphonic midi out. If you want the midi though I'd guess cleaning up the delayed mess wouldn't be too crazy provided you know the delay time.

Good luck

[osiris]

FL now has a Deverb function in the Audio Clip menu. Any body tried it?