Zero EQ

AI-assisted project. This codebase was created with Claude (Anthropic), directed and reviewed by a human author. The DSP has been verified analytically (filter math checked against the RBJ/Butterworth cookbook formulas) and against real audio (throwaway numeric test harnesses processing real signals through the actual shipped processor class, not just static curve math — including FFT-verified harmonic content, generated per-band, for the Harmonic character), pluginval passes clean on VST3 and passes on AU with two known benign warnings (see Status), and it's been loaded and hosted successfully in REAPER. It has not been used on real hardware in a live signal chain yet. Review before use on live gear.

A zero-added-latency parametric EQ + compressor VST3/AU plugin, built with JUCE.

Real screenshot of the Standalone build.

Goals

• Zero added latency — every filter is minimum-phase IIR (biquads); no lookahead, no linear-phase FFT convolution, no oversampling, no internal buffering beyond the host's block size. Safe to track through in a live/monitoring signal chain.
• Node-based curve workflow — draggable per-band nodes over a live pre/post spectrum analyzer. Drag = frequency/gain, scroll = Q, double-click = add/toggle a band.
• Dynamic EQ bands — any gain-having band (Bell/Shelf/Tilt) can be switched into dynamic mode: it gets its own threshold/ratio/attack/release/range and behaves as a frequency-selective compressor (duck when that band gets loud) or expander (boost when that band gets quiet), instead of sitting at a fixed static gain. Detection runs on the signal already in the chain by default — no lookahead, no added latency — or, per band, on an external sidechain input instead.
• Three per-band characters — Modern (independent Q, textbook response), Vintage (proportional Q that widens with applied gain, approximating passive/console-style musical EQs such as Cranborne Audio's Harmonic EQ), and Harmonic (generates gain-driven even/odd harmonics from only that band's frequency region and sums them back in parallel with the dry signal, so it adds harmonic content where the band is aimed instead of saturating the full-range signal). Vintage and Harmonic are original approximations of that behaviour, not circuit models or clones of any specific product.
• Input/output trim with metering, plus a post-EQ feed-forward compressor (soft-knee, peak/RMS detection, no lookahead — also zero added latency).
• Presets — 7 factory presets covering every feature area (static EQ shaping, dynamic EQ, harmonic generation, compressor-forward), plus save/load of your own presets to the standard per-user preset directory. Factory presets double as the plugin's VST3/AU host "program" list.
• Ballistics-accurate metering — input/output meters combine a fast peak read, a slower VU-style integrating level, a 4x-oversampled true-peak estimate that catches inter-sample peaks a plain sample read would miss, and a held clip indicator, all on a read-only side path that can't add latency to the signal actually being processed.

Roadmap

• Dynamic EQ bands — shipped. Per-band threshold/ratio/attack/release/range, selectable Downward (duck) or Upward (boost) direction.
• Harmonic-based EQ — shipped. A third per-band character that generates gain-driven even/odd harmonics from only the band's own frequency region and sums them back in parallel, with a per-band blend control between the two harmonic families.
• Preset system — shipped. 7 factory presets plus user preset save/load.
• Ballistics-accurate metering — shipped. Peak, VU, true-peak-estimate, and held clip indication on the input/output meters, replacing the earlier raw peak-only reads.
• External sidechain input for dynamic EQ — shipped. An optional stereo sidechain bus, with a per-band toggle so any dynamic band can detect against that external signal instead of the internal chain.

All five landed without compromising the zero-added-latency guarantee that's the whole point of this plugin. Next up: whatever the next real driver of this project turns out to be — nothing currently queued.

Signal chain

flowchart LR
    IN["Input Gain<br/>+ metering"] --> EQ["8-band EQ<br/>minimum-phase biquads<br/>+ per-band dynamics<br/>+ per-band harmonic generation"]
    EQ --> COMP["Feed-forward compressor<br/>soft-knee, no lookahead"]
    COMP --> OUT["Output Gain<br/>+ metering"]

Loading

Each EQ band's own dynamic detector reads the signal as it arrives at that band's position in the chain (post every earlier band, pre this one) and modulates that band's gain in real time. The Harmonic character then runs a small parallel branch off that same point: the band's frequency region is split out, shaped, and only the harmonics that shaping generated are summed back into the main path — so a Harmonic band colours the frequencies it targets and leaves the rest of the spectrum alone. Still just one pass through the chain, no lookahead anywhere.

Dynamic EQ visual feedback

Dynamic bands get three layered visual cues on the main curve so you can read what's actually happening without opening the band panel:

• Range envelope — a soft shaded region showing the full swing a dynamic band could reach (static gain out to its configured Range, in its Downward/Upward direction). Purely a function of the band's own settings, so it's visible even on silence.
• Live curve — a second curve, drawn on top of the static one, showing the actual instantaneous total gain (static + live dynamic delta) across the whole spectrum. It visibly moves in real time as the signal pushes each dynamic band's detector, with the gap between it and the static curve filled in to make the current deviation obvious at a glance.
• Node engagement glow — each dynamic band's node ring brightens and thickens in proportion to how hard it's currently working (current delta relative to its Range), so an idle dynamic band reads as a faint outline and an actively-ducking/boosting one visibly glows.

Building

Requires CMake 3.22+ and a C++20 compiler (Xcode Command Line Tools on macOS). JUCE is fetched automatically via CMake FetchContent on first configure.

cmake -B build -G Ninja -DCMAKE_BUILD_TYPE=Release
cmake --build build

Build products (VST3 / AU / Standalone) land in build/ZeroEQ_artefacts/.

EQ bands

Each of the 8 bands supports: Bell, Low Shelf, High Shelf, High Pass, Low Pass, Notch, Band Pass, and Tilt Shelf, with a selectable slope (12/24/36/48 dB/oct) for the HP/LP types and a Modern/Vintage/Harmonic character switch for the rest.

Bell/Shelf/Tilt bands can additionally be switched into dynamic mode: threshold, ratio, attack, release, and a max-range clamp, with a Downward (duck) or Upward (boost) direction. The detector filters a copy of the signal through a type-appropriate analysis shape (band-pass at the band's freq/Q for Bell/Tilt, high-pass at the corner for High Shelf, low-pass at the corner for Low Shelf) to isolate "the region this band affects," then envelope-follows that to drive the gain modulation. This is a practical approximation for isolating a band's spectral region, not a claim of matching any specific commercial dynamic EQ's exact detection algorithm.

Each dynamic band also has its own Sidechain toggle: leave it off and the band detects against the signal already in the chain (the default); turn it on and it detects against the plugin's dedicated stereo sidechain input bus instead — useful for, say, ducking a bed track off a separate voice input. The toggle falls back to internal detection automatically (rather than detecting against silence) whenever the host hasn't actually connected the sidechain bus, and its label switches to "Sidechain (n/c)" in that case so it's obvious from the GUI. The sidechain bus adds no channels to, and never writes to, the main signal path — it's a read-only detection input, so it doesn't touch the zero-added-latency guarantee either.

The Harmonic character blends between two waveshapers driven by how much gain the band is applying (a band left at 0dB stays transparent even in Harmonic mode): an asymmetric quadratic shaper that generates warm, tube-like even harmonics (plus a DC blocker, since asymmetric shaping introduces a DC offset), and a tanh shaper that generates grittier, transistor-like odd harmonics. A per-band blend slider crossfades between them. Both use first-order antiderivative anti-aliasing (ADAA) rather than oversampling — the waveshaper's antiderivative is evaluated across the current and previous sample instead of the raw function at the current sample, which suppresses aliasing from the nonlinearity without needing lookahead or extra latency.

Crucially, the shaper is fed only that band's spectral region, not the full-range signal. The band's audio is split off through the same region-isolation filter the dynamic detector uses (band-pass at freq/Q for Bell/Tilt, high-pass at the corner for High Shelf, low-pass for Low Shelf), saturated, and then only the generated harmonic content — the shaped signal minus the isolated region — is summed back into the dry path. That parallel structure is what makes this a harmonic EQ rather than a saturator behind a filter: a Harmonic bell at 5 kHz adds harmonics derived from the 5 kHz region and leaves a bass note passing through the same band untouched, and because only the difference is summed, the band's own level is unchanged.

Note that drive is coupled to the band's gain (|gain| / 12, clamped) rather than being an independent control — a Harmonic band at 0dB is exactly transparent, and you reach full drive at ±12dB. That keeps the character honest (no harmonics without an EQ move) but it does mean you can't currently dial in heavy harmonics at a modest boost. Roughly, measured on a bell at Q=1 with the even-leaning blend, the 2nd harmonic sits about 28dB below the fundamental at +3dB gain, 15dB below at +6dB, and 6dB below at +9dB; +12dB is deliberately extreme.

Presets

7 factory presets, each starting from a full reset-to-defaults so every parameter not mentioned below stays at its default:

Preset	Showcases
Init	Flat/default state, useful as a starting point or reference
Vocal Presence	Static EQ shaping — low-mid cut, presence bump, air shelf
De-Esser (Dynamic)	A single dynamic downward band tuned to catch sibilance
Warm Bus (Harmonic)	Harmonic character on two bands (even-leaning low end, odd-leaning top) plus gentle bus compression — subtle by design, both bands sit at low gain and therefore low drive
Podcast Voice	HPF + presence + a dynamic band taming harshness + compressor, combined
Broadcast Loudness	Compressor-forward: higher ratio, faster attack/release, auto makeup
Telephone / Lo-Fi	Aggressive band-limiting plus an odd-leaning Harmonic band driven around mid-range, for creative use

User presets save to the standard per-user preset directory (~/Library/Audio/Presets/Allan Sargeant/Zero EQ/ on macOS) as plain APVTS-state XML — the same save/restore path a host already uses for session recall, so there's no separate preset format to keep in sync. Factory presets also drive the plugin's VST3/AU "program" list, so a host's own program-switching UI works too.

Metering

Input and output each get their own read-only LevelMeter instance sitting after that stage's gain trim — it reads a copy of the buffer and never touches the actual signal, so nothing it does can add latency:

• Peak — instant attack, ~20dB/s exponential release. A fast-peak-meter approximation, not a claim of matching a specific broadcast PPM's exact spec.
• VU — a one-pole RMS-based integrator tuned for ~300ms to reach 99% of a step, approximating classic VU ballistics (not a model of a real meter's needle physics).
• True peak — a lightweight 4x-oversampled estimate using Catmull-Rom cubic interpolation between samples, catching inter-sample peaks a plain sample-peak read misses. A practical estimate, not a full ITU-R BS.1770-compliant filter.
• Clip indicator — held for ~1.5s so a brief clip stays visible, triggered at a small headroom margin below exact 0dBFS rather than an exact >= 1.0 check (a nominally unity-gain stage isn't always bit-exact 1.0 after a normalized-parameter round-trip, so a razor-edge threshold can miss a genuinely full-scale sample).

Unsigned builds — macOS Gatekeeper & Windows SmartScreen

The release plugins and standalone app are not code-signed or notarized — that needs paid Apple / Windows developer certificates this project doesn't carry. They're safe to run; the OS just can't verify a publisher. On macOS this matters more than usual: a quarantined plugin will make your DAW skip it or fail validation rather than show a friendly prompt.

macOS

After copying the bundles into place, clear the quarantine flag:

xattr -dr com.apple.quarantine "$HOME/Library/Audio/Plug-Ins/VST3/Zero EQ.vst3"
xattr -dr com.apple.quarantine "$HOME/Library/Audio/Plug-Ins/Components/Zero EQ.component"

(Use /Library/... instead of $HOME/Library/... if you installed them system-wide.) Then rescan plugins in your DAW. If the AU still won't appear, the Audio Unit cache may be stale — quit the host and run killall -9 AudioComponentRegistrar, or just reboot. For the standalone app: right-click → Open the first time, or xattr -dr com.apple.quarantine "Zero EQ.app".

Windows

Windows doesn't gate .vst3 plugins the way it gates .exe files, so your DAW loads Zero EQ.vst3 normally. The standalone .exe may trip SmartScreen — More info → Run anyway.

Linux

No signing gate — copy Zero EQ.vst3 into ~/.vst3/ and rescan.

Signing it yourself (optional)

macOS ad-hoc signing stops the host from re-quarantining on your own machine (it is not notarization — it won't clear Gatekeeper on someone else's Mac):

codesign --force --deep --sign - "Zero EQ.vst3"
codesign --force --deep --sign - "Zero EQ.component"

Distributing without warnings needs an Apple Developer Program membership ($99/yr) + a Developer ID Application certificate — sign each bundle with the hardened runtime (codesign --options runtime) and notarize the containing .zip/.dmg with xcrun notarytool, then xcrun stapler staple the .dmg.

Status

Phase 7: DSP engine, dynamic EQ (with optional external sidechain detection), band-isolated harmonic generation, presets, ballistics-accurate metering, and full interactive GUI (spectrum analyzer, draggable curve, preset bar, band/compressor/IO panels, live dynamic-gain indicators). Verified via pluginval (VST3 + AU, strictness 5 — VST3 clean; AU passes with two known benign warnings, see below), hosted successfully in REAPER, and checked against real audio through the actual shipped processor class — including a two-tone FFT check confirming the Harmonic character generates harmonics from only its own band's region while leaving an out-of-band tone untouched (91dB of selectivity measured, with the out-of-band fundamental passing through within 0.01dB), that pure-odd blend produces odd harmonics and no even ones, and that a Harmonic band at 0dB gain is fully transparent; a full sweep confirming every factory preset applies correctly, produces finite (no NaN) audio and reports zero added latency; a metering pass confirming peak attack/release timing, VU ballistics, true-peak inter-sample detection, and clip-indicator hold time all match their intended behaviour on known test signals; and a sidechain pass confirming a dynamic band tracks the external sidechain signal (not the internal one) when its toggle is on, tracks the internal signal steadily when it's off, and falls back to internal detection cleanly when sidechain audio isn't actually connected. See open items below for what's still outstanding.

Known benign AU warnings: two pluginval warnings appear consistently on the AU build while every other check — and Apple's own auval validator, exit code 0 — passes clean:

• "Disabling non-main buses failed" — a documented JUCE/AudioUnit interaction. AUv2's bus-disable semantics for auxiliary/sidechain input buses don't round-trip the way pluginval expects, independent of what the plugin does; a widely reported quirk for AU plugins carrying a sidechain bus.
• "Current program is -1" — the AU wrapper reports no factory preset as explicitly selected until a host selects one, even though getCurrentProgram() returns a valid index and the factory preset list is fully populated and functional.

Known limitations / next steps

• Not yet tested against real (non-silent) audio hardware in a live signal chain.
• No linear-phase mode — intentionally out of scope (zero latency was the explicit goal).