Why Suno and Udio Stems Turn Harsh Under a Limiter and How to Clean Them Without Dulling the Mix

Your stems sound fine in the DAW until the moment a limiter touches the master bus, and then the vocal and the cymbals turn into tinsel. That thin, glassy, slightly metallic edge is the sound most people mean when they talk about AI fizz, and it is the single most common reason a Suno track that felt finished in the browser suddenly sounds cheap on headphones.

Here is the short version. The harshness is not spread across the whole top end. It lives in a handful of narrow ringing bands, and the limiter drags those bands forward because it raises everything quiet while squashing everything loud. A wide high shelf cut kills the fizz and the real air together. Narrow dynamic cuts on the actual problem spots, applied per stem before anything hits the bus, remove the fizz and keep the brightness. And a simple phase null test, where you listen only to what you are removing, tells you in seconds whether you have gone too far.

I have cleaned a lot of these stems. Some of what follows is measurement and some is my own ears and mistakes, and I will be clear about which is which.

Why the Limiter Makes It Worse Instead of Hiding It

A limiter does two things at once. It stops peaks from crossing a ceiling, and in the process it brings up the average level, which means everything that was sitting politely underneath the loud parts gets lifted. Transients get flattened, the gap between the loud snare hit and the quiet ring behind it shrinks, and detail that used to be masked is suddenly audible.

AI fizz is exactly that kind of detail. On a raw stem it is sitting a long way below the vocal or the cymbal. Once the limiter compresses the difference, the ratio between the real sound and the ringing artifact gets smaller, and the artifact moves to the front. It was always there. The limiter just turned up the lights.

This is also why the problem feels like it appears out of nowhere. People check their stems soloed at a moderate level, decide they are clean, and only hear the glaze when they push for loudness at the end, by which point the fizz from five or six stems is stacking up on the bus.

Where the Fizz Actually Lives

The breakthrough for me was looking instead of just listening. Open a decent spectrum analyzer on a harsh vocal stem, slow the averaging down, and watch the top end while the fizz is audible. You will not see a broad lump above 15 kHz. You will see a few thin, stubborn vertical streaks that stay lit regardless of what note is being sung, plus a scattering of narrower spikes higher up.

In the stems I have measured, the trouble clusters in two regions. There is a persistent ring in the upper mids, somewhere around 1.4 kHz on a lot of vocal stems, and a cluster of thin spikes between roughly 14 and 18 kHz on vocals and cymbals alike. The exact positions move a little from song to song and from model version to model version, so treat those numbers as where to start looking rather than where to put your cuts blindly.

The reason this happens comes down to how these stems are made. As several engineers pointed out in a long VI-Control discussion about Suno audio quality, Suno and Udio generate the whole song as one mixdown, then split that mixdown into stems afterwards. They are separation outputs, not the original tracks. Separation leaves thin spectral residue behind, and the generation model itself leaves its own smeared, slightly watery texture on the top end. RoEx, who build mixing tools and have written a detailed breakdown of why Suno stems sound bad, describe the highs in these stems as inconsistent and note that phase relationships between stems are not always coherent either. That phase point matters later.

Harshness that lives in narrow bands needs narrow tools. Brightness that lives everywhere else needs to be left alone.

The High Shelf Mistake Almost Everyone Makes First

The advice you will see most often is to roll off the top end with a shelf or a low pass filter somewhere around 15 kHz. I did this for months. It works, in the sense that the fizz goes away. It also removes the breath on the vocal, the shimmer on the hats, and most of what makes a mix feel open. The result sounds like a blanket over the speakers, and people then try to add the air back with an exciter, which brings the fizz straight back.

The shelf is the wrong shape for the problem. If the artifact were a broad hiss, a shelf would be right. The artifact is a lattice of thin ringing lines, so a shelf is cutting ninety percent good signal to reach ten percent bad signal. A narrow notch or a narrow dynamic band does the opposite.

There is a second reason the shelf fails. The upper mid ring near 1.4 kHz is nowhere near 15 kHz, so a top end shelf never touches it, and that ring is a big part of why AI vocals sound boxy under compression.

Dubspot’s guide to mixing AI-generated stems recommends hunting for problem frequencies between 2 and 5 kHz on harsh vocals and using a de-esser between 5 and 8 kHz, which is sensible general advice. I would add the lower ring and the very top cluster to that hunt, and I would do the hunting with a delta test rather than a sweep.

Listen to What You Are Removing

This is the one technique I would keep if I had to throw everything else away. Instead of judging a cut by listening to the processed stem, listen only to the difference between the processed stem and the original. If that difference is nothing but a thin whistle, you have found the artifact and nothing else. If you can hear the singer’s words or the snap of the snare in the difference, you are cutting real music and need to back off.

Many plugins have this built in. FabFilter Pro-Q 4 and oeksound’s soothe3 both have a delta or difference mode. If your EQ does not, you can set it up by hand in any DAW in about a minute. Duplicate the stem onto a second track. Put your cleanup EQ on the duplicate only. Flip the polarity on the original, which is the little phase button that most DAWs put on the channel strip or in a utility plugin. Send both tracks to a bus and listen. Everything the two tracks have in common cancels, and what remains is exactly what your EQ changed.

A few things keep the test honest. Both tracks need to be sample aligned, so do not nudge either one. Match the output gain of the EQ so you are not hearing a plain level difference. Linear phase mode on the cleanup EQ makes the null cleaner, because minimum phase EQ shifts phase around the cut and leaves a faint smear of the whole track in the difference that is not a sign of over cutting.

With the difference signal running, the cuts become easy to judge. A narrow band at 1.4 kHz should sound like a faint flute tone in the null, and a cluster of cuts above 14 kHz like a thin, airy hiss with no musical content. The moment the null sounds like a quieter version of the song, you have gone too far.

Static Notch, Dynamic EQ, or Resonance Suppressor

There are three ways to make these narrow cuts, and they are not interchangeable. Which one you use depends on the stem and on how much the fizz moves around.

A static notch is a fixed, narrow cut at one frequency. It is the simplest option and works for an artifact that rings at the same pitch all the time, which is often true of the very top cluster on cymbal stems. The downside is that it cuts even when there is no fizz to cut, so on vocals it leaves a small hole in the tone during quiet passages.

A dynamic EQ band only cuts when the level in that band crosses a threshold. On a vocal stem the fizz flares up on loud sustained notes and is almost inaudible in soft phrases, so a dynamic band bites on the loud notes and gets out of the way the rest of the time. This is the approach I use on nearly every AI vocal now. It is also where I noticed the biggest difference on my own tracks, because the glassy sheen on an AI vocal sits in the same high range as the cymbal junk and a static cut dulls the whole voice.

A resonance suppressor, which is the category soothe3 belongs to, listens for whatever is ringing above the surrounding spectrum and pulls it down automatically, wherever it happens to be. It is the laziest option and often the best one on badly mangled stems, because the artifact spikes move slightly as the model hallucinates its way through a phrase and a fixed band cannot follow them. The risk is that pushed too hard it flattens the tone and starts to sound processed in its own way.

Here is how the main options line up on price and fit, with prices as listed by each developer or by the plugin guides I checked while researching this article.

TDR Nova deserves a word on its own. It is free, it does proper dynamic EQ, and the Sample Focus roundup of the best EQ plugins of 2026 calls it their favorite free dynamic equalizer, with the only real limit being four bands in the free version. For a lot of people reading this, Nova is the whole answer and the paid plugins are a later luxury.

Pro-Q 4 at $199 earns its price if you are doing this every week. The delta mode, the spectrum display you can grab bands from, and the linear phase option make the whole workflow faster. Dubspot’s head to head of Pro-Q 4 and TDR Nova puts it fairly: Pro-Q buys speed and refinement, Nova buys capability for nothing.

soothe3 is $259 in the US, or 229 euros and 199 pounds in the EU and UK, sold directly by oeksound. According to Sound On Sound’s coverage of the soothe3 launch in May 2026, the new version is more aware of the overall tonal balance of the source than soothe2 was, which means a wider sweet spot and fewer new problems introduced while fixing old ones. MusicTech’s review notes it ships with delta controls so you can hear exactly what it is pulling. It is expensive for a single processor, and if you only clean a stem or two a month, Nova will get you most of the way there for nothing.

Settings That Work on Vocals Versus Cymbals

These are starting points from my own sessions. Use the null test to confirm every one of them on your stem rather than copying them blindly.

For vocal stems, start with a dynamic band near 1.4 kHz with a Q between 8 and 12. A Q value describes how narrow a band is, and higher numbers are narrower. Set the band to cut up to 4 or 5 dB, then raise the threshold until it only moves on the loud sustained notes. Add two or three narrow dynamic bands in the 14 to 18 kHz region, each cutting up to 6 dB, with the same loud notes only threshold. Attack fast, release around 80 to 150 milliseconds so the band lets go before the next syllable.

Vocals need a lighter touch than drums. Push the high bands hard and you start grabbing the S sounds, and the vocal gets lispy. When that happens, back the maximum cut on the high bands down to 3 dB and let a dedicated de-esser handle sibilance between 5 and 8 kHz, placed after the cleanup EQ and before any compression.

For cymbal and hat stems, you can be more aggressive. Static notches work at the very top because the ringing on cymbal stems tends to stay put. Cut 6 to 9 dB at each spike with a Q of 15 or higher. The upper mid ring is usually less of an issue on drums, so skip the 1.4 kHz band unless the null test tells you otherwise. Cymbals are the easiest win, and the improvement is usually obvious on the first pass.

For instrument stems like guitars, keys, and synths, check them but do not assume they need work. Many come through clean enough that any cut is a loss. If you find yourself adding narrow notches on the master, go back and find which stem is responsible.

Do the Cleanup Per Stem, Before the Bus

This is where a lot of careful people go wrong. They mix the stems and then reach for a mastering chain to fix the harshness at the end. Mastering cannot do this job well, because the fizz on six stems sits at six slightly different frequencies, and a narrow cut on the bus either misses most of them or has to be so wide it dulls everything. Jack Righteous makes the same point in his guide on reducing shimmer and hiss in AI music: mastering is a finishing layer for a mix that already works, not a repair layer for a damaged source.

So the order is clean each stem, then balance, then compress on the bus, then limit. Put the cleanup EQ at the very start of each stem’s chain. Compression evens out the level difference the dynamic band relies on, and saturation generates new harmonics on top of the ringing, which makes the artifact louder. Clean first, then shape.

The phase issue RoEx flagged is worth a check at this stage too. Because the stems were pulled from a mixdown by a separation model, summing them back together does not always reproduce the original mix cleanly, particularly in the low end. Flip the polarity on the bass stem and listen to the kick. If the low end suddenly gets fuller, leave it flipped. Separation models occasionally hand you a stem that is partially inverted against its neighbors.

Get the Best Stems You Can Before You Clean Anything

Cleanup is easier when the source is better, and the way you export matters more than people think.

On Suno, there are now two different stem paths. The quick route is the Get Stems or Auto Split option on a track in your library, which is available on the Pro plan at $10 a month or $96 a year and splits a song into up to 12 time aligned WAV files. The other route is Suno Studio, which is Premier only at $30 a month or $288 a year. According to Dubspot’s report on Suno Studio 2.0, which shipped on August 13, 2026, Premier subscribers can export multitracks and stems from Studio at 32 bit and 48 kHz with no cap, while the regular library downloads have been capped since September 3, 2026 at 20 songs a month on Pro and 60 on Premier. A separate review from aireiter notes that the menu extraction delivers 16 bit files, so if you are on Premier and care about cleanup, export from Studio.

Always take WAV, never MP3. Lossy encoding adds its own smearing in the same 14 to 18 kHz region where the fizz already lives, and once the two are mixed together you cannot separate them.

Newer models help too. Suno moved to v6 on September 9, 2026, and the stem splitter got quality updates in June 2026, so regenerating an older song on the current model is worth a try. In my experience the fizz does not vanish, but it gets thinner, which makes every cut less invasive. The separation step will keep leaving residue for as long as stems are derived from a mixdown.

Udio is a different story in late 2026. As Billboard reported, Udio disabled all downloads on October 29, 2025 as part of its settlement with Universal Music Group, briefly reopened them for a 48 hour window in November, and has kept them off since while it builds a licensed platform that keeps creations inside the app. According to Dupple’s September 2026 comparison of Suno and Udio, which checked both vendors’ pricing pages directly, there are currently no audio, stem, or video downloads on any Udio plan, and Udio’s terms limit output to personal and non commercial use. Udio stems saved before that date can be cleaned exactly as described here. New ones cannot be pulled into a DAW right now, and anyone selling a workaround is selling you a problem.

When Cleanup Stops Being the Answer

There is a point past which no amount of surgical EQ saves a stem. Narrow cuts can remove ringing that sits on top of a sound, but they cannot restore detail that was never there. If a cymbal stem is a smeared wash with no real transient, cutting the fizz leaves you with a quieter smeared wash.

For drums, the practical fix is replacement. Keep the AI kick and snare if they are usable, and swap the hats and cymbals for samples or a short recorded pass. Hats and cymbals are the stems most likely to be unsalvageable and the cheapest to replace, and a real cymbal under an AI vocal does more for believability than any plugin.

For vocals, the honest answer is sometimes to re-record. An AI vocal with heavy glaze that still sounds synthetic after cleanup is telling you the problem is in the source. Having a real singer re-cut the vocal over the cleaned instrumental, or having real musicians rebuild the whole arrangement from the AI track as a reference, is one legitimate route among several. It also solves the ownership question that pure AI output still carries, since a human performance is something you can clearly register and release. It costs more than a plugin, and it is not the right call for every song, but for a track you actually care about it is often the step that finally makes it sound like a record.

A middle path is to re-amp an AI guitar part through a real amp, or use Suno Studio’s MIDI export on Premier to drive a real synth, which gives you a clean part that keeps the original idea.

Checking the Final Master Without Fooling Yourself

Once the stems are clean and the mix is balanced, the master is where you find out whether it worked. A few checks save a lot of grief.

Listen at the loudness you intend to release at, not quieter. Spotify’s own loudness normalization documentation says the platform plays tracks at minus 14 LUFS integrated, which is a measurement of average loudness, and recommends keeping true peaks below minus 1 dBTP, or below minus 2 dBTP if your master is louder than minus 14 LUFS, to avoid distortion in the encoder. LUFS stands for loudness units relative to full scale, and dBTP means decibels true peak. If you master hotter than that, which plenty of people do, Spotify will just turn you down on playback, so the only thing extra loudness buys you is more fizz exposure for no gain.

Run the null test one last time on the full mix. Bounce a version with all the stem cleanup bypassed, line it up against the clean version, flip the polarity, and listen to the difference. It should sound like a thin, airy whistle with no recognizable song in it. If you can hum the melody from the null, something upstream is cutting music.

Listen on a phone speaker too. The 1.4 kHz region is exactly where small speakers have the most energy, so an upper mid ring that is barely noticeable on monitors can be the loudest thing in the mix on a phone.

Finally, compare against a commercial reference in the same genre at matched loudness. The reference keeps you honest about how much top end a real record actually has.

A Pre-Flight Routine You Can Run This Week

Here is how I handle a fresh batch of stems now. It takes about twenty minutes per song once it becomes habit.

Export every stem as WAV at the highest bit depth your plan allows and drop them into a fresh session. Solo each stem, push it through a temporary limiter on its own channel, and listen for glaze. That mimics what the bus limiter will do later and tells you which stems need work, usually the vocals and anything with cymbals. Leave the clean stems alone.

On each stem that needs work, load TDR Nova or Pro-Q 4 first in the chain, find the thin vertical streaks on the analyzer, place a narrow dynamic band on each one, and adjust in the delta view until the difference signal is pure fizz with no song in it. Then build the rest of the chain behind the cleanup, balance the mix, check bass polarity, and only then add the bus compressor and limiter. If glaze appears at release loudness, go back to the stem causing it rather than reaching for a shelf on the master.

If a stem fails all of that, replace it. A sample for the hats, a re-recorded vocal, or a real musician rebuilding the part is not a defeat. It is the same decision a producer would make with a badly recorded live track, and it is the step that separates a Suno stems cleanup that sounds decent from one that sounds like a finished record.

Sources

• RoEx, Suno Stems Sound Bad? Here Is Why and How to Fix Them: https://www.roexaudio.com/blog/suno-stems-sound-bad-here-is-why-and-how-to-fix-them
• Dubspot, How to Mix AI-Generated Stems: A Practical Cleanup and Mixing Guide: https://blog.dubspot.com/mixing-ai-generated-stems-2026
• VI-Control, SUNO audio quality vs sample libraries: https://vi-control.net/community/threads/suno-audio-quality-vs-sample-libraries.172526/
• Jack Righteous, Reduce AI Music Noise: Fix Shimmer, Hiss and Generation Artifacts: https://jackrighteous.com/en-gb/blogs/guides-using-suno-ai-music-creation/reduce-ai-music-noise-fix-shimmer-hiss-in-suno-ai
• Dubspot, Suno Studio 2.0 Adds MIDI, a Synth, and Unlimited Stem Export: https://blog.dubspot.com/suno-studio-2-0
• aireiter, Suno Studio 2.0 Review: MIDI, Wavetable Synth, Stem Export Math: https://aireiter.com/blog/suno-studio-2-0-review
• DIYAI, Suno Pricing 2026: Plans, Credits and Commercial Use: https://diyai.io/ai-tools/audio-generation/suno-pricing/
• Dupple, Suno vs Udio 2026: Which AI Music Generator Is Better?: https://dupple.com/learn/suno-vs-udio
• Billboard, Udio Says Users Can Download AI Songs for 48 Hours After Backlash to UMG Legal Settlement: https://billboard.com/pro/udio-deal-backlash-ai-users-download-ai-songs-48-hours
• Sample Focus, Best EQ Plugins of 2026: https://samplefocus.com/blog/best-eq-plugins-of-2026
• Dubspot, FabFilter Pro-Q 4 vs TDR Nova: https://blog.dubspot.com/plugins/compare/fabfilter-pro-q-4-vs-tdr-nova
• Sound On Sound, oeksound introduce Soothe3: https://www.soundonsound.com/news/oeksound-introduce-soothe3
• MusicTech, Is Oeksound’s Soothe3 the most transparent resonance suppressor yet?: https://musictech.com/reviews/plug-ins/oeksound-soothe3-review/
• Spotify for Artists, Loudness normalization on Spotify: https://support.spotify.com/au/artists/article/loudness-normalization/