Shimmer

by The Treq

Clean up AI music artifacts. Then master for release.

A free, offline mastering suite built for tracks made with Suno, Udio, and other AI music tools.

---

The problem

You generate a track and it sounds great — until you really listen.

There's a thin, fizzy sizzle riding on top of the cymbals. The vocals have a glassy, plastic sheen. The high end sounds busy in a way that real recordings don't. Turn it up on good speakers and it gets worse.

That noise is a generation artifact. AI music models build audio in ways that leave behind hiss, flicker, and metallic ringing in the high frequencies — usually between 5 kHz and 12 kHz. It is not in your melody, your mix, or your performance. It was added by the model.

Normal mastering tools can't fix it. They treat that noise as part of your music, so they compress it, brighten it, and make it louder.

What Shimmer does

Shimmer does two jobs, in the right order:

1. Fixes what you hear. You pick what bothers you from a set of cards under What do you hear? Each card turns on one tool made for that problem: a notch filter for a steady whistle, a de-esser for harsh "s" sounds, and so on. You set how hard each one works.
2. Masters the result. A loudness target, tone shaping, and true-peak limiting, so your song is ready to upload.

That order matters. Fix first, master second. If you master first, you just make the fizz louder.

New in 2.0

Shimmer 2.0.0 is a rebuild of 1.1.1. The screens look much the same. The engine behind them is new.

• Cards instead of presets. The 19 presets and the nine-stage cleaner are gone. You pick what you hear, and each card turns on one tool that was tested on its own.
• One sound path. The preview, the export, Batch, album mode, the Remix export, and the command line all use the same function, render(). What you hear in the preview is what the file will be.
• Louder by default. The default loudness target is now Commercial (−9 LUFS). In 1.1.1 it was Streaming (−14 LUFS), so masters came out quiet next to released songs.
• A tighter limiter. It finds peaks at 8× oversampling (1.1.1 used 4×) and aims just under the ceiling, so peaks no longer slip past it.
• Safer MP3, OGG and M4A files. 1.1.1 used −1.5 dBTP for every lossy format, and a decoded M4A could reach +1.24 dBTP. Now lossy formats get −2.0 dBTP, and every lossy file is decoded and checked after encoding.
• Filters land on their setting. In 1.1.1, preset shelves and bells landed at twice their setting: −3 dB became −6 dB.
• OGG export works. In 1.1.1 every OGG export failed.
• Your old settings carry over. A saved 1.x preset turns on the card it became.

Every change that alters the sound, with its numbers, is in docs/SOUND-CHANGES.md. What 2.0 does not do yet is listed under Known limits.

---

Hear exactly what you're removing

This is the part most cleanup tools don't give you.

Shimmer lets you listen to three versions of your track, and switch between them instantly with the 1 2 3 keys while the song plays:

Key	Track	What you hear
1	Original	Your untouched upload
2	Processed	Fixed and mastered
3	Removed	Only what the fixes took out — turned up so you can hear it clearly

That third one is your safety check. The Removed track should sound like the problem you picked: hiss, fizz, a whistle, harsh "s" sounds. If you hear vocals, snare hits, or melody in there, a fix is working too hard. Turn its Amount down.

Because volume fools your ears, Loudness-matched A/B is on by default. It evens out the levels between versions, so you judge the sound instead of just picking whichever is louder. The bar tells you what it is doing, for example "Processed −2.2 dB". That is a monitoring level only. Your export is never touched.

The transport at the bottom has back-to-start, skip 5 seconds either way, play/pause, and a scrubber you can click or drag. The amber band on the scrubber is the Live loop window.

---

Quick start

New to Shimmer, or to mastering? Follow the step-by-step Getting started guide.

Windows

1. Download the latest release → grab Source code (zip), then unzip it somewhere permanent (not your Downloads folder).
2. Double-click start.bat.
3. Wait. Your browser opens on its own when Shimmer is ready. Drop a track in and go.

What happens on that first launch: Shimmer needs a free helper called uv to install itself. If you don't have it, start.bat offers to install it for you — just press Enter. It then builds its own private Python environment and downloads the audio libraries (up to 200 MB). After an update, it installs again if the list of libraries changed.

Budget 5–10 minutes for the first run. After that, Shimmer starts in a few seconds. Nothing is installed system-wide. The app's libraries go in a .venv folder next to the app, and deleting the Shimmer folder removes them. Your saved settings are kept in their own small folder (%APPDATA%\Shimmer on Windows).

Windows may warn you about running a downloaded file. Click More info → Run anyway. You can read every line of start.bat in a text editor first — it's plain text with a comment on each step.

macOS / Linux

git clone https://github.com/henricksmedia/shimmer.git
cd shimmer
./start.sh

(Or download the latest release, unzip, and run ./start.sh from that folder.)

That's it. start.sh does the same thing as the Windows launcher: it sets up uv if you don't have it (asking first), builds a private environment, installs the audio libraries, then opens your browser once the server is actually ready.

Same expectation as Windows — 5–10 minutes the first time, seconds after that. The libraries go in a .venv folder next to the app; delete the Shimmer folder to uninstall. Saved settings are in ~/.config/shimmer.

Prefer to do it by hand?

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
python -m uvicorn shimmer.server:app --port 7860

Then open http://localhost:7860.

Everything works the same as on Windows, with two differences worth knowing:

• Stem separation runs on the CPU. GPU acceleration currently looks for an NVIDIA card, so on Apple Silicon the Remix tab still works — it's just slower. Results are cached per track, so you only wait once.
• The Batch tab's "Browse…" buttons need Tk. Most Python installs include it. If the picker doesn't open, type or paste the folder path into the box instead — that works everywhere. (Homebrew users: brew install python-tk.)

Windows is the most-tested platform simply because that's what it was built on. If you hit a macOS or Linux issue, please open an issue — bug reports from other platforms are genuinely useful.

Updating from 1.1.1

Get 2.3.5 the same way you got 1.1.1 (a new release zip, or git pull in your clone), then launch it as before.

• A git clone keeps itself up to date. From now on, start.bat and start.sh check GitHub each time they start, and move a clone on the main branch forward to the newest release. They leave a clone alone when it has unsaved changes or no network, and a zip download is never touched. To try a branch without that, use start-test.bat (or start-test.sh): it runs that copy as it is, on port 7870, with its own settings.

• The libraries update themselves. start.bat (Windows) and start.sh (macOS and Linux) see that the library list changed and install what 2.0 needs before they start.

• Your saved settings carry over. A 1.x preset turns on the card it became. For example, Suno Hash turns on Shimmer, and Muddy / Boxy turns on Low-mid build-up. python -m shimmer --list prints the whole table.

• Stems you already split are kept, as long as they are in the same Shimmer folder. A song's fingerprint did not change, so the stem cache and saved Remix projects still match.

One optional extra

WAV, FLAC, and OGG work right away. MP3 and M4A also need ffmpeg:

Windows	winget install ffmpeg
macOS	brew install ffmpeg
Linux	apt install ffmpeg

If something goes wrong

• "uv was not found" → say yes to the install prompt, or install it yourself (winget install --id=astral-sh.uv -e on Windows, brew install uv on macOS), then launch again.
• The window closes instantly → open a terminal in the Shimmer folder and run the launcher from there so you can read the error.
• macOS: "permission denied" → run chmod +x start.sh, then ./start.sh again. Or just run bash start.sh.
• Browser opens but the page won't load → give it another few seconds and refresh. The first start is the slow one.
• Still stuck? Open an issue and paste what the black window says.

---

Your first master, in 3 steps

1. Drop your track in. WAV, MP3, FLAC, OGG, or M4A.

2. Click Analyze, then pick what you hear. Analyze measures your song. It looks for fixed tones and turns on the Fixed tones card if it finds any. It tells you how far the song is under your loudness target. It moves the Live loop to where the top end is busiest, and it plans a short Suggested EQ. It takes about 10 seconds for a four-minute song, and it never changes your audio.

Analyze only reports what it can measure well today. For the other cards, your ears decide. Listen, then turn on the cards that match what you hear. Each card you turn on gets an Amount slider. Start where it starts, check the Removed track, and adjust.

3. Click Clean & Master. A progress window lights up each stage as it runs, then closes when your finished file is ready — along with a What changed chart (the spectrum before and after, and what was removed) and the numbers behind it: loudness before and after, true peak, peak-to-loudness ratio, stereo correlation, and how hard the limiter worked. A Release check card gives one verdict on the file, "Ready to upload" or what to look at: loudness on target, true peak under the ceiling, clipping in the upload, sample rate and format, silence at the start and end, length, DC offset, mono compatibility, the bass in mono, and tags. Then it shows how loud the song will play on Spotify, Apple Music, YouTube and the rest. The window ends on a Download step. In Settings you can have the file download by itself, or land straight in a folder of yours with a Show in folder button.

Want to hear your edits right away? Turn on Live in the bottom bar. It loops a short part of your song and renders it again after each change. The loop is the same render as the export, cut to the loop's length, so what you hear is what the file will be. A test holds the two together: the difference between them must be at least 60 dB quieter than the song.

---

What's inside

"What do you hear?" — ten cards, one tool each

Pick the cards that match what you hear. Each one turns on the tool made for that problem.

Card	What it sounds like	Tool	Status
Shimmer	Fizzy, flickering hiss up top	Spectral de-noise	On, to try
Vocal grain	Grainy hiss riding on the voice	Voice de-noise	New in 2.0.1, to try; picked by ear, its Amount top not yet measured
Fixed tones	A whistle or whine that never changes	Notch filter	Ready; Analyze turns it on
Sibilance	Harsh, spitty "s" and "sh"	De-esser	On, to try
Clicks and crackle	Short pops, ticks or static	De-click	Not built yet
Harshness	Piercing, painful upper mids	Dynamic EQ (2–5 kHz)	On, to try
Phasiness	Grainy or watery reverb tails	—	No fix yet
Low-mid build-up	Muddy, boxy, words hard to hear	Dynamic EQ (200–500 Hz)	On, to try
Lack of air	Dull, no sparkle	Tone target (in Mastering)	Ready
Loudness	Quieter than released music	Loudness target (in Mastering)	Ready

"On, to try" means the tool passed its measurements but has not had its blind listening round yet. A card marked "Not built yet" or "No fix yet" changes nothing. You can still mark it, and Shimmer counts those picks on your computer to help test new fixes.

Each fix has an Amount slider. Its top is set by measurement: at 100 %, a fix takes no more than a small, fixed amount from a clean song (0.10 sones on a hearing model). The tests behind each fix are in docs/STEP6-FIXES.md.

Shimmer — spectral de-noise. A small trained model sets a gain for each frequency between 1.5 and 16 kHz. It runs on your computer, in numpy, so nothing extra is installed and nothing is uploaded. The first time the card is on for a song, it reads the whole song once: about 50 seconds for a 3-minute song. The screen shows how far it has got, and turning the card off stops it. After that, a new Amount or a new preview takes about a third of a second.

Fixed tones — notch filter. AI generators often leave a few thin, steady tones at one pitch for the whole song. Analyze finds them across the whole file and turns the card on. Narrow notch filters cut each one, on both channels, at full depth. On the test models the notch removes 96 % of a fixed tone.

Sibilance — de-esser. Turns the high band down only while an "s", "sh", "t" or "ch" sticks out: fully in the center, half as much in the sides. Up to 6.5 dB at Amount 100 %.

Harshness and Low-mid build-up — dynamic EQ. Finds the band that sticks out most (up to two in 2–5 kHz for Harshness, one in 200–500 Hz for Low-mid build-up). It cuts that band only while it gets louder than usual for that song. Up to 3.3 dB for Harshness and 4.0 dB for Low-mid build-up at Amount 100 %.

Lack of air and Loudness are handled in Mastering. Loudness sets the loudness target to Commercial (−9 LUFS). Lack of air brightens the top end through the tone target.

Mastering that respects your dynamics

• Loudness targets. LUFS is how streaming platforms measure loudness.

  Target	LUFS	What it means
  Commercial (default)	−9	As loud as most released songs
  Balanced	−11	A little quieter, with more punch left in
  Streaming standard	−14	The level streaming apps play songs at; sounds quiet in other players

• One steady gain. Loudness is set with one gain, worked out from the whole song. There is no multiband compression. A 25 Hz low-cut (high-pass filter) comes first. It runs one way, so it adds no pre-echo before a kick.

• Peak shaper and true-peak limiter. The shaper, a soft clipper at 4× oversampling with both channels linked, rounds off the peaks. Then the limiter holds the ceiling. It finds peaks at 8× oversampling and eases the gain down across a 2 ms lookahead, so it does not click.

• Tone target. With mastering on, Shimmer moves your song's tone toward a target: its built-in one, or a reference track you pick. Tone match (Low, Medium, High) sets how much. Tone (warmer to brightest) tilts it. The built-in curve boosts at most +5 dB and cuts at most −3 dB, and it never boosts above the point where your render's top end stops.

• Reference-track matching. Pick a released song you like as the tone target. Shimmer compares the two tone shapes, level-matched, and moves your song part of the way: 50 % of the difference by default, smoothed over about an octave, and never more than ±3 dB. The Master tab shows both shapes and the curve it will apply. It warns you when the reference has far more or far fewer drum hits than your song.

Export formats

Format	Details	True-peak ceiling
WAV 24-bit (default)	Your song's own sample rate	−1.0 dBTP
WAV 16-bit 44.1 kHz	The release copy, with TPDF dither	−1.0 dBTP
FLAC	24-bit	−1.0 dBTP
FLAC 16-bit 44.1 kHz	The same audio as the WAV release copy, about half the size	−1.0 dBTP
MP3	320 kbps, 48 kHz at most	−2.0 dBTP
OGG Vorbis	Quality 0.8	−2.0 dBTP
M4A (AAC)	256 kbps	−2.0 dBTP

16-bit files get TPDF dither: a very quiet, even noise, one step either way, so quiet fades do not turn grainy. Codecs push peaks up when they encode, so every MP3, OGG and M4A is decoded after encoding. If one would go over −1.0 dBTP, it is turned down by the excess and encoded again, and the report says by how much.

A real EQ, when you want one

A 12-band parametric EQ with a curve you can drag: bells, low and high shelves, high-pass, low-pass, and notch filters. Drag a point to move it, scroll on it to change the Q (how wide or narrow the move is), double-click to add or delete.

Bells and shelves land exactly on their setting. The high-pass and low-pass run one way, so they add no pre-echo. Thirteen starting points are built in, at mastering scale (small, broad moves): corrective ones like Rumble cut, Tighten lows, De-mud, Box cut and Smooth the top; tonal ones like Tilt darker, Tilt brighter, Warmth, Open mids, Presence, a gentle Air lift, and Vocal clarity; and Lo-fi telephone for when you want an effect. Every band a starting point loads stays editable.

Suggested EQ: the moves a mastering engineer would make

Analyze also plans a short EQ for the track, the way a pro works: fix first, shape second, cuts before boosts, nothing big. It listens to the loud parts only. It looks for a ringing tone, a build-up in the low mids, and a harsh band in the presence range. Then it makes at most three gentle balance moves toward a family range — Neutral, Pop, Hip-hop, EDM, Rock, R&B, Acoustic, Lo-fi, or Cinematic. A family is a tolerance, not a target: it says how far from neutral your track may be before a move is worth making.

It judges the song the way it will render: after the fixes and, with mastering on, after the tone curve, so nothing is corrected twice. Every plan is checked on the loudest 20 seconds. If the peaks would rise more than the loudness, the limiter would work harder, so boosts are halved, then dropped. You see the moves and the reasons, you can turn any move off, and Apply puts them into the EQ as normal bands you can edit.

Tags that travel with the file

Exports carry proper metadata: title, artist, album, genre, year, track, copyright and ISRC, written the way each format expects (WAV gets both RIFF INFO and an ID3 chunk, so Windows Explorer shows them too). The file's own tags stay. Shimmer fills the blanks from your defaults and adds a note to the comment, so a file always says what was done to it. Set your artist name once in the Tags section and every export has it.

Remix: split the song into stems

Break a finished track into vocals, drums, bass, and other (or six stems with guitar and piano), then mix the parts in a lane mixer, one row per stem with its own waveform:

• Mute, solo, fader and pan on every lane
• Formant — shift vocal character without changing the pitch
• Saturation — warmth and drive
• Doubler — thickens a part like double-tracking it
• Reverb — room and depth
• A Residual lane holding whatever the separator dropped, so the untouched mix is your original, with a null-test figure that says how much is in it
• Quick mixes: instrumental, acapella, vocal lift

Pick a quality: Fast (about 9 seconds for a 4½-minute song on an RTX 4070 SUPER), Best (the fine-tuned model, about 22 seconds), 6 stems, Ultra (models averaged with extra passes, about 4.5× Best), or Studio (a Mel-Band RoFormer vocal model, then Best for the rest). Loop it, export a remix, or download the stems themselves as a ZIP of 24-bit WAVs. Stems are kept per track and quality, so you only wait for a split once. Your mix saves itself automatically.

The Remix export runs through the same render and export as the Master tab. The loop preview plays at once from the loop's own mix. A few seconds later it switches to that same render, and the status line says it matches the export.

Separation uses your GPU when you have one. The first run downloads the separation engine, which is a large one-time install; the Best model fetches another 330 MB the first time you pick it.

Models and licences. Shimmer ships no separation model. Each split downloads its model the first time you pick it, from the author's own release, into stem_cache/ next to the app, and works offline after that. A model file you place in stem_cache/models/ yourself (same file name) is used as it is. Every model offered may be used on music you release and sell:

Split	Model	Author	Licence
Fast, Best, 6 stems, Ultra	htdemucs, htdemucs_ft, htdemucs_6s, hdemucs_mmi	Meta Platforms (Demucs)	MIT
Studio	Mel-Band RoFormer vocal model, then htdemucs_ft	Kimberley Jensen; Meta	MIT

Models whose authors have published no terms are not offered. The full list with links is in NOTICE, and each card in the Remix tab names its model, author and licence.

Batch: a whole folder at once

Point Shimmer at a folder and let it work. Every file goes through the Master tab's own render and export, with tags and the release check. With Auto-detect on, each file gets what Analyze finds (Fixed tones today), and the log lists what it found in each file. Results stream in file by file as it goes, with each file's loudness and true peak when mastering is on.

Album mode masters the folder as one record. Shimmer measures every track, then picks one gain that brings the loudest track to the target. The others keep their distance below it, so a quiet song stays quieter than the single instead of every track being pushed to the same level. The log shows the loudest track, the gain, and the album's overall loudness.

Batch can also plan a Suggested EQ for each file on its own, and write your Tags defaults onto every export.

The Signal Chain tab

A map of every stage your audio passes through, in order, drawn from the settings you have right now: Read, Trim, Sample rate, Fixes, Tone, EQ, Master, Export, and Report. Each stage says whether it will run and, when it won't, why. The map is built from the same rules render() follows, so it shows what a run will actually do.

---

How it works

Every tab sends your song through one function, render(), in this order:

1. Sample rate. If the format needs a different rate (the 16-bit release copies are 44.1 kHz), the song is resampled first. That way the limiter's ceiling holds at the rate that is written.
2. Fixes. One tool per card that is on: the notch filter first, then the spectral de-noise, the de-esser, and the dynamic EQ. Each tool works out its plan once from the whole song, so a preview window gets the same treatment as the export.
3. Tone. With mastering on, the tone curve moves the song toward the built-in target or your reference track.
4. EQ. Your parametric EQ.
5. Level. With mastering on: the 25 Hz low-cut, one loudness gain worked out from the whole song, the peak shaper, and the true-peak limiter. With mastering off and Preserve volume on: one gain that puts the song back at its own level.
6. Export. Dither for 16-bit files, encoding, the lossy check and tags. Then the release check reads the file that was written.

Analyze, the Suggested EQ planner, and the Signal Chain tab only measure. They never change the sound.

---

Known limits

2.0.0 is the first release of the new engine. These are known, and each is planned for a later version:

• The Remix and Batch tabs still show the old preset menu. Behind the scenes, each preset turns on the card it maps to. The cards come to both tabs later.
• Clicks and crackle and Phasiness have no working fix yet. The de-click is built, but it misses moderate pops in dense music, so the card says "Not built yet". Phasiness has no fix at all yet.
• The fixes marked "On, to try" have not had their blind listening rounds yet. The Shimmer fix removes the flicker it was trained on. On the test models, it does little for the other kinds of fizz.
• All four fixes at full Amount can take a little too much. With Shimmer, Sibilance, Harshness and Low-mid build-up all at 100 %, two of the five test songs lost a little more than one fix is allowed to take. 2.0.1 brought Sibilance and Harshness back under that limit; more sound tuning is planned.
• Batch has no cancel button.
• Icons show as words when you are offline. The icon font loads from Google Fonts.

---

Your music stays on your computer

Shimmer runs entirely on your machine. Your tracks are never uploaded, never sent to a server, and never used to train anything. There are no accounts and no subscriptions. After setup, it works without an internet connection. The only things it fetches are the page's fonts (when you are online) and a stem model the first time you pick it.

---

FAQ

Will this make my track sound dull? It can if you push it too far. That's what the Removed track is for — listen to it. If you hear real music in there, lower that card's Amount. Every Amount slider already stops where its fix starts to take too much from a clean song.

Do I need to know what LUFS or true peak means? No. Leave the target on Commercial and Shimmer handles it. The numbers are shown for people who want them.

Which cards should I turn on? Click Analyze first. It turns on Fixed tones when it finds any, and it tells you if the song is quieter than your target. Then listen, and turn on the cards that match what you hear. Check the Removed track after each one.

What happened to my preset? Presets are gone in 2.0. Your saved preset turns on the card it became, at that card's starting Amount. python -m shimmer --list shows which preset became which card.

Does the Shimmer fix send my music anywhere? No. It is a small trained model that ships with Shimmer and runs on your computer, in numpy. Nothing is uploaded and nothing extra is installed.

Can I use this on regular (non-AI) recordings? Yes, though it's tuned for AI artifacts. The mastering, the EQ, and stem remixing work on any audio.

Is it really free? Yes. Free to download, free to use, and free to use on music you sell. The licence only matters if you want to copy, change or share Shimmer's code, or run it as a service: that needs a written licence. See License below.

Can I use it on tracks I'm selling? Yes. The licence covers the software, not your music. Anything you make with Shimmer is yours, with no strings and no royalties.

---

For developers

Command line, for scripting and batch jobs:

python -m shimmer input.wav output.wav --target cd
python -m shimmer input.wav output.wav --fix tones=0.5 --no-auto
python -m shimmer input.wav release.wav --master --release
python -m shimmer --suggest input.mp3
python -m shimmer --list

The command line runs the same render and export as the Master tab. --list shows the cards, the loudness targets, the formats, and which 1.x preset turns on which card. Mastering stays off unless --master or --target asks for it, as in 1.x. The old 1.x cleaning flags are accepted and ignored, with a note. Run python -m shimmer --help for every option.

Run the test suite:

python -m pytest tests/

Where to read more:

• docs/README.md — the developer overview: the sound path, the project layout, and the tests.
• docs/ARCHITECTURE.md — how 1.x worked, why it was rebuilt, and the plan and rules for 2.0.
• docs/API.md — every HTTP route the screens call.
• docs/SOUND-CHANGES.md — every sound change against 1.1.1, with numbers.
• docs/STEP6-FIXES.md — how each fix was measured.
• docs/REBUILD-TRACKER.md — where each step of the rebuild stands.
• CHANGELOG.md — version history.

Architecture in brief: FastAPI backend, plain JavaScript frontend (no build step), NumPy and SciPy for the sound work. The engine is shimmer/core/, the web routes are shimmer/api/, shimmer/server.py serves the app, and shimmer/cli.py is the command line.

---

Contributing

Fixes, new tools for the cards, and platform improvements are all welcome — especially macOS and Linux bug reports, since Shimmer was built on Windows. Open an issue or a pull request.

By sending a change, fix or idea, you let Henricks Media use, change and license it as part of Shimmer, with no payment owed (section 4 of the licence).

License & credits

Shimmer is released under the Shimmer License. It is not open source: the code is public so you can read it, not so you can reuse it.

In plain language:

• ✅ Use it free, for anything — including music you sell. Anything you make with Shimmer is yours. The licence covers the code, never your music.
• ✅ Read the code to see how it works, and report bugs.
• ❌ Copying the code, changing it, sharing it, hosting it as a service, or using any part of it in other software needs a written licence from Henricks Media.

Versions before 2.2.0 were released under the AGPL-3.0, and copies of those versions stay under it.

The Shimmer and The Treq names and logos belong to Henricks Media.

Copyright, trademark, and third-party component licences are listed in NOTICE.

Built on excellent open-source work: NumPy and SciPy for the DSP math, FastAPI for the server, pyloudnorm for loudness measurement, soundfile for audio I/O, and Demucs (Meta, MIT) with audio-separator (MIT, carrying RoFormer model code by lucidrains) on PyTorch for stem separation. Each keeps its own license.

The separation models download on first use from their authors' releases and stay on your machine: the Demucs models by Meta (MIT) and the Mel-Band RoFormer vocal model by Kimberley Jensen (MIT). Shimmer only offers models whose license allows use on music you release and sell; the full list, with each license, is in NOTICE.

---

Built by an AI music creator, for AI music creators

I'm not a trained musician — I make music with these same AI tools. Shimmer exists because I got tired of hearing that fizz on my own tracks, and you shouldn't need an audio engineering background to fix it.

🎵 Hear my music — The Treq

_{© 2026 Jeremy Henricks · Henricks Media · All rights reserved · Shimmer License}