Shimmer

The problem

Music generated with tools like Suno and Udio has a characteristic flaw. There’s a thin, fizzy sizzle riding on the cymbals and a glassy sheen on the vocals, concentrated roughly between 5 and 12 kHz. It isn’t in the melody or the performance — the model put it there.

Conventional mastering makes it worse. A normal chain treats that noise as part of the music, so it compresses it, brightens it, and makes it louder. The order of operations is the actual insight: clean first, master second. Master first and you’ve just amplified the fizz.

Why no machine learning

The cleanup path contains no ML at all. It’s classical DSP, and that’s a deliberate choice: the same settings produce the same result every time.

For a tool that sits between someone’s creative work and its release, that predictability is worth more than a marginally better average result. A neural restoration model that occasionally invents detail or behaves differently across runs is a bad trade when the user is trying to make a release decision. Demucs handles stem separation, where a learned model genuinely is the right tool — but it stays out of the artifact path.

The surgical approach

The design assumption is that most of the track is fine and shouldn’t be touched.

Low frequencies are never processed. A linear-phase crossover splits at 4.5 kHz. Kick, bass and vocal body bypass the cleaning engine entirely and rejoin untouched.

The center of the mix is protected. Above the crossover, audio splits into mid and side. Vocals and snare live mostly in the center, so that channel is cleaned at roughly 20% strength, while the sides — where most generation shimmer lives — get full treatment.

Nine detectors run on the high band, each targeting a different artifact shape: noise floor, resonance notching, core shimmer suppression, harshness, flicker, grid patterns, whistles. Two safety gates ride along — one backs off during noisy or percussive passages, another protects transients for 70 ms so drums keep their attack.

Then it masters: high-pass, a single steady loudness gain, soft clip, and a 4× oversampled true-peak limiter. Loudness is one gain move rather than multiband compression, so the dynamics survive.

The feature I care about most

Shimmer lets you audition three versions of the track and switch between them with the number keys while it plays: the original, the processed result, and only what was removed, turned up so it’s clearly audible.

That third one is the honest one. The removed track should sound like hiss and fizz. If you hear vocals, snare or melody in it, the tool is cutting too hard and you should back it off. Most cleanup software asks you to trust it; this one hands you the evidence to check it.

Loudness-matched A/B comparison is on by default, because volume fools ears and otherwise people just pick whichever version is louder.

Also built in

• 19 artifact presets grouped by what the user actually hears — “glassy cymbals,” “plastic vocals,” “muddy” — rather than by DSP parameter
• Analysis that evaluates a track against every preset, picks the best fit, explains why in plain language, and locates the worst passage
• Live preview that loops a section and re-renders in about a second
• 12-band parametric EQ, zero-phase so it shapes tone without smearing transients, with a draggable curve
• Stem remixing into vocals, drums, bass and other, with per-stem formant shifting, saturation, doubling and reverb
• Batch processing across a folder, with optional per-track auto-detection
• Signal chain view explaining every stage in order, in plain language

Product decisions

It runs entirely offline. No accounts, no uploads, no telemetry — someone’s unreleased music never leaves their machine, which for this audience is a requirement rather than a feature.

The launcher builds its own isolated Python environment on first run and tells the user to expect 5–10 minutes, because the honest version of a slow first-run experience is to say so. Deleting the folder uninstalls it completely.

It’s AGPL-3.0 with a commercial license available, has a CLI alongside the web interface, and runs pytest in CI. Documentation covers macOS and Linux caveats directly, including CPU-only stem separation on Apple Silicon.

Why it’s here

I’m not a trained audio engineer, and the README says so. I built this because I make music with these tools and got tired of the fizz.

It’s in this portfolio because it’s the clearest evidence of something specific: taking an unfamiliar technical domain, learning it properly enough to make sound architectural decisions — deterministic over clever, protect what’s working, show the user your evidence — and shipping a documented, tested, licensed tool that other people can actually run.