AI-based removal of aliasing post-recording

[Makrophag]

Hi,
I did a search but could not really find anything.
Main method to remove aliasing is to filter audio before coverting to digital or use oversampling.

However no method seems to exist up to now, to reduce/remove aliasing from existing recordings. Right so, because it is impossible for DSP to identify the aliasing noise from the source audio.
I wonder, if an AI could be trained to learn to remove this?

(I came to this because I got vocal stems that were massively effected using a sample rate reduction on purpose. I would love to get the clean voice back)

Best

[Jorchime]

Neat idea. I think this could work since aliasing typically has some sort of pattern to it.
You'll still get artifacts, similar to the stem-separator AIs etc.
But this could become a fix-tool for scenarios where you can't reproduce the original without aliasing.

With filtering and oversampling before printing something to file, we have great tools to avoid it altogether, though. So I'd really see this just as a niche tool to fix rare issues.

[stian]

Hi,
I did a search but could not really find anything.
Main method to remove aliasing is to filter audio before coverting to digital or use oversampling.

However no method seems to exist up to now, to reduce/remove aliasing from existing recordings. Right so, because it is impossible for DSP to identify the aliasing noise from the source audio.
I wonder, if an AI could be trained to learn to remove this?

(I came to this because I got vocal stems that were massively effected using a sample rate reduction on purpose. I would love to get the clean voice back)

Best

Thanks for the suggestion! I've had the same thought, and it would be an interesting experiment. I'm not sure if aliasing is still a problem that many users struggle with, but from an academic standpoint, it would be very interesting.

Best,
Stian

[Makrophag]

Yes, I agree.

Probably, in this type of source material (low quality digital audio), bit depth also plays a role.
Thinking if this, one application might be working with sounds from old video games.

[Tj Shredder]

Hi,
I did a search but could not really find anything.
Main method to remove aliasing is to filter audio before coverting to digital or use oversampling.

However no method seems to exist up to now, to reduce/remove aliasing from existing recordings. Right so, because it is impossible for DSP to identify the aliasing noise from the source audio.
I wonder, if an AI could be trained to learn to remove this?

(I came to this because I got vocal stems that were massively effected using a sample rate reduction on purpose. I would love to get the clean voice back)

Best

Thanks for the suggestion! I've had the same thought, and it would be an interesting experiment. I'm not sure if aliasing is still a problem that many users struggle with, but from an academic standpoint, it would be very interesting.

Best,
Stian

It would be worthwhile for restoration tasks. I bet there are a lot of early digital recordings which could be repaired that way…
Training could be done with several converter models…