If you had a single stereo A.I. track and wanted it to sound less harsh and more finished without stems or re synthesis,what DSP steps would you try first . Thanks in advance for any input
The first DSP step is arguably the most difficult one, but I’d put my money on the 4th one, especially when it comes to all things harshness and finishes.
I’m sorry to say I think this is the wrong forum for such a question.
I would suggest you learn the basics of EQ (equalization and filtering) to tame harshness. You do not need to program anything custom for this, you can use plenty of free tools for processing a stereo audio file.
To make it sound more “finished”… Well that’s a whole subjective term in itself. But if you don’t want to learn anything about mastering, you can use something like Landr or iZotope’s Ozone. Or you can hire a human engineer who will produce much better results more reliably (citation: https://www.youtube.com/watch?v=wZRV2H4PK0Q).
If you want to learn these skills, it’s a long road, but starting off with some mixing and mastering tutorials will get you there.
I don’t think you need/want to learn C++ and the JUCE framework, unless you want to build a plugin or app for this, but I would advise against that unless you already have the tools and expertise necessary.
Thanks for taking the time to reply.i should clarify im not looking to learn mastering or build a juce plug in. Im exploring the possibility of whether there is room for a lightweight,post generation process aimed at casual creators who like their track but experience fatiguing across common listening platforms. Thanks for your reply.
So just a quick follow up on my previous post. I’ve now passed identical lyric sets through multiple Suno generations, using a couple of different style targets, tempo changes etc. What I’m finding consistent isn’t the music, but the failure mode.. I’m seeing fast delivery and dense consonant clusters seem to collapse into the same kind of smearing and slurring, regardless of style and speed. It feels to me, less like bad vocals and more the systems hitting a resolution ceiling when it’s exposed to pressure.
I’d be interested to hear if anybody has seen this kind of behaviour at the waveform / phonetic level?. Many thanks in advance
Quick update on my previous posts..
After much listening . The one thing that has emerged. Is it possible the real issue sits more in how the vocals are being forced to relate to a highly regular grid, rather than the generator itself?. Again, many thanks in advance
Just an update.
I’m not approaching this as any kind of expert in DSP or audio systems. I have a strong conviction,built through hours of repeated listening,testing and failure, that there’s perhaps something here worth closer examination. Interested to understand how people with the relevant experience would frame it, test or even dismiss the idea.
Does it sit within the realms of practcal feasibility ? I’m not here to try and define possible solutions, but to create conditions where expertise could engage with the question. Thanks for the read and any future input.. kind regards
Since you are trying to fix something generated by AI to begin with, maybe just throw more AI into it to fix the results of the other AI? 
There is no general easy automatic way to make things sounds “better”, “less fatiguing” etc. Those are such subjective, vague and broad things. Or if someone has figured out a good way, they’d of course keep that a secret, sell that as a product and become rich. 
Hey.. thanks for the reply..I’m not actually looking for a general or automatic “make it sound better”solution…im looking to answer the narrower question ,whether constrained micro-timing irregularities within the transitions can measurably reduce perception of rigidity and fatigue in a.i . generated audio .. if its “no” or “already well understood”.. that’s a perfectly valid outcome.. many thanks for your response. Appreciate.
If you are looking for more controls over the vocal, perhaps Dreamtonics Synthesizer V or ACE Studio would be a better choice. They give you much more control over the vocals so that you can fine-tune each word
Disclaimer: I am not affiliated and I haven’t used them for about two years.
1 Like
Hey.. thanks. I’m aware of tools that allow deep per word control. I’m exploring if its possible that some of the micro-intervention can be somewhat reduced, upstream, so the first pass could, perhaps ,feel less rigid, pre manual refinement.. thanks for your response 
Hi all,
Following up from my earlier posts, I’m looking at whether tiny, controlled micro-timing variations at phrase or onset boundaries in a single stereo AI-generated track can measurably reduce perceived rigidity or listening fatigue — essentially a pre-humanisation step before any manual edits.
Constraints:
No stems, no re-synthesis
Adjustments limited to temporal deviations that preserve content, harmonic balance, and timbre
I’d greatly value insights on DSP approaches, ways to model or measure perceptual effects, or any prior work in this area. Even a sanity check on feasibility would be very helpful.
Thanks in advance for any pointers or discussion.
To clarify scope before I step back from this thread:
I’m not looking for downstream mastering, DAW workflows, or per-event manual “humanisation”.
The narrow question I’m interested in is whether perceived rigidity and listening fatigue in single stereo AI-generated renders can plausibly be linked to overly deterministic temporal structure at micro-transition boundaries (e.g. inter-phoneme, transient-to-sustain, or phrase junctions), and whether small, bounded, non-random deviations applied upstream or immediately post-render are known to measurably affect that perception.
In other words: is there any established DSP or psychoacoustic basis for introducing constrained micro-timing or envelope irregularities (as opposed to noise, modulation, or EQ) to reduce perceived rigidity in otherwise static material — or is this effect already well understood to be negligible or ill-posed?
A clear “no, this is not a meaningful DSP problem” would be just as useful as pointers to relevant literature or prior art.
Follow-up: perceptual DSP intervention on finished stereo mixes
A few months ago I asked about improving listening comfort in AI-generated music using post-render DSP. Since then I’ve narrowed the scope considerably.
The question is no longer about “fixing” vocals or source separation. I’m only interested in whether a small perceptual improvement band exists when working strictly on a finished stereo mix.
Constraints:
Input: rendered stereo track with embedded vocals
No stems available
No source separation or ML reconstruction
Processing must be lightweight deterministic DSP
No structural remixing or editing
The observation prompting the question is that many AI-generated tracks exhibit upper-band vocal harshness / spectral edge that creates listening fatigue even when the musical content is otherwise acceptable.
Core question:
When operating only on the final stereo signal, is there a meaningful perceptual improvement band where constrained DSP (e.g. dynamic spectral shaping, de-harshing, masking-aware processing) can reduce perceived harshness or fatigue without introducing obvious artefacts?
I’m not expecting a universal solution—just curious whether developers working on audio effects have encountered similar limits or techniques when dealing with mixed stereo material.
Interested in any thoughts from people who build or work on perceptual audio processors.
adding low-pass filter at 20Hz would produce the most natural result
1 Like
Thanks for sharing the link. The video seems to focus on automated mastering services, which isn’t quite the problem space I’m looking at.
I’m not trying to automate mastering or improve mixes in a general sense. The question I’m exploring is narrower: whether small perceptual improvements can be made on a finished stereo render using lightweight DSP (e.g., dynamic spectral shaping) to reduce upper-band harshness or fatigue in AI-generated material when stems aren’t available.
So the interest is really about the limits of post-render DSP, rather than mastering automation.
I don’t think you’ll find any help in these forums. Most members are against AI-generated content. Not only is your problem not interesting to them, but the fact that every single post you have written is also generated by AI is a big turn-off.
Nobody here is interested in solving AI-related problems. You let a plagiarism machine generate garbage and then wonder how to improve the output to be less garbage.
At that point, it’s too late.
2 Likes