Lancet — user manual

Cut where it counts — a surgical dynamic EQ with an analog soul.

What's new in v0.4.0

The Attack knob is now true down to 0.1 ms. Up to v0.3.0 there was a fixed 50 ms smoother sitting behind the gain computer, so no Attack setting faster than about 50 ms could actually be heard - 0.1 ms, 1 ms and 20 ms all did the same thing. That smoother is gone; the detector's own ballistics are now the only thing shaping how fast a band moves, evaluated once per sample.

Please read this if you have existing sessions. A session using a non-zero Range with a fast Attack will react faster than it did before. This is a bug fix, not a re-voicing - the plugin now does what its own knob always said it did - but it is a real, audible change and you should expect to hear it. If a band now feels too grabby, dial its Attack up: for the first time, that control has the effect its label describes. Sessions with Range = 0 on every band (pure static EQ, no dynamics) are unchanged, and the static EQ curves themselves are identical to v0.3.0's in every band type, gain and Q.

Also new:

• Sidechain (SC Source: Internal / External). Each band can now take its detector input from an external sidechain instead of the audio passing through it - so a band can duck against a kick, a lead vocal, or anything else your host can route. Enable the plugin's sidechain input in your host first; if there is no sidechain available, a band set to External simply keeps using the internal signal rather than going quiet.
• SC Mode: Split / Wide. Split (the default, and the only behaviour previous versions had) means the band only listens to its own frequency region. Wide means it listens to the whole signal while still only moving its own band - the usual way to make one band pump with the mix. Listen follows whichever you pick, so you always hear what is really triggering the band.
• Cleaner Saturation. The per-band Saturation stage now uses an anti-aliased waveshaper. Same character, noticeably less of the gritty fold-back harmonics that a plain waveshaper produces on high-frequency content - and still no added latency.
• Smoother Gain/Q. With Gain/Q coupling on, the band's Q now glides with the dynamic gain instead of stepping.
• An eleventh factory preset, Sidechain Carve, demonstrating the new routing.

Both new per-band controls sit in the editor as combo boxes under each band's Type slot. Sessions saved by older versions load exactly as before, with both new controls at the settings that reproduce the old behaviour.

What's new in v0.3.0

A musical-voicing pass (see docs/voicing-notes.md) - measured where the plugin's own DSP made a measurement possible, honestly labeled "by ear, not yet tuned against real material" where it doesn't:

• Per-band Q/Threshold/Attack/Release defaults, tuned to each band's documented role (was a flat Q 1.0/Threshold -30 dB/Attack 5 ms/ Release 150 ms for every band): Band 1 (100 Hz, boom/sub control) now starts slow and gentle (Attack 25 ms/Release 280 ms); Band 5 (4 kHz, sibilance/harshness) starts fast (Attack 2 ms/Release 70 ms); the bands in between step down progressively. Range still starts at 0 dB (idle) for every band regardless - nothing moves until you dial in a Range, but once you do, each band now reacts the way its role suggests it should. See the updated per-band table below.
• Gentle Saturation (bN_sat, new per-band toggle, off by default): when on, a soft tanh-based waveshaper is applied to a band, but only while it's actively boosting (Gain + Range's dynamic contribution net positive) - a cutting or idle band is completely unaffected even with Saturation on. Automation/preset controllable in v0.3.0; a dedicated editor toggle is roadmap M3, same as Auto Release/Gain-Q.
• A tenth factory preset, "Analog Warmth Lift" (docs/presets.md): demonstrates the new Saturation toggle on a gentle low-mid boost.
• A v0.2.0 session loads cleanly into v0.3.0 (tolerant import): every existing parameter value is preserved exactly, and the new per-band Saturation toggle populates at its off default.

What's new in v0.2.0

A research-derived deep-dive rework (see docs/design-brief.md/ docs/research-notes.md), the suite's M2 preset system, and a German frame localisation:

• Program-dependent Auto Release (bN_autoRelease, new per-band toggle, off by default): when on, a band's effective release time shortens automatically - never slower than the manual Release setting - when the signal's own envelope is already falling on its own (e.g. a naturally decaying transient), inspired by (not a reproduction of) the "ARC"-style release behaviour documented in F6-class dynamic EQs. Automation/preset controllable in v0.2.0; a dedicated editor toggle is roadmap M3.
• Gain/Q coupling (bN_gainQ, new per-band toggle, off by default): when on, a band's own filter Q softens (widens) proportionally to how hard its dynamic gain is currently moving, for a gentler, more analog-style character at deeper dynamic moves - the band's static Gain never affects Q, only its dynamic component does. Same automation/preset-only status as Auto Release in v0.2.0.
• Attack/Release ranges widened: Attack now 0.1-500 ms (was 0.5-100 ms), Release now 5-1500 ms (was 10-1000 ms) - both ends, reaching both faster transient-catching and slower, musical tonal-balancing use cases.
• Knee width is now derived from Range, not a flat 6 dB constant - shallower Range settings now read as gentler/softer, full-depth (±12 dB) Range settings sound identical to v0.1.0's fixed 6 dB knee.
• Eleven factory presets (docs/presets.md) covering common use cases (glue, de-essing, transient enhancement, mix-buss settling, slow tonal balancing, resonance taming, and a diagnostic Auto Release demo), plus a preset bar (Save/Save As/Delete/Import/Export, factory + user library) at the top of the editor.
• A v0.1.0 session loads cleanly into v0.2.0 (tolerant import): every existing parameter value is preserved exactly, and the two new per-band toggles populate at their off default.

What it is

Lancet is a six-band dynamic EQ in the spirit of the Waves F6 class - cited here as a documented reference point for the category, without implying endorsement, sponsorship, or affiliation by Waves Audio Ltd. Each band is a normal parametric EQ band (bell, or shelf on Band 1/Band 6) whose gain can additionally move with the program material. Feed it loud and it reacts — cutting a resonance only when it flares up, or opening a boost only when a part gets buried — then settles back to its static setting once the signal drops back down. Because each band's dynamic move is driven by its own pre-EQ, band-filtered detector, one band's cut never confuses another band's detector, and a band's own gain move never feeds back into its own detection.

Where a static EQ band asks "how much?", Lancet's dynamic bands also ask "when?" — the difference between permanently notching out a 3 kHz resonance (which also thins the tone whenever that resonance isn't present) and only pulling it back exactly when it rings.

Where it sits in a mix chain

Lancet is a corrective, surgical tool, most useful early-to-mid signal chain, before broad tonal shaping and bus compression:

Source track -> [gain staging / gate] -> Lancet (resonance/harshness control) -> broad EQ / saturation -> compression -> bus

Reach for it when a static EQ cut would either under-treat the problem (leaving room for it to still poke through on the loudest hits) or over-treat it (thinning the tone on quieter passages where the problem isn't present). It also works as a mix-bus or master-bus tool for controlling a specific recurring resonance or harshness band without permanently coloring everything under it.

Signal flow

in --[Input Trim]--+--[pre-chain tap]--> each band's Detector (bandpass @ band freq/Q -> envelope)
                    |
                    +--> Band1 -> Band2 -> Band3 -> Band4 -> Band5 -> Band6 --> [Mix] --> [Output Trim] --> out

Every band's detector taps the signal right after Input Trim, before Band 1 - not that band's own serially-processed input - so a downstream band's gain move never perturbs an upstream band's detection, and no band's own move feeds back into triggering itself. See docs/architecture.md for the full engineering breakdown (gain-computer formula, detector selectivity, sub-block coefficient smoothing, Listen).

Under the hood

The per-sample gain path and the TPT SVF core

Up to v0.3.0 a band's realised gain came from a 50 ms juce::SmoothedValue sitting after the gain computer - the detector could react in half a millisecond and the filter would still take 50 ms to catch up, so most of the shipped 0.1-500 ms Attack range did nothing audible. v0.4.0 deletes that smoother and evaluates the whole chain - detector envelope, dB conversion, soft-knee gain computer, Range clamp, filter gain - once per sample. The detector envelope is the smoother now, and there is no stepped-gain fallback for any band type.

Deleting the smoother from a coefficient-rebuilding biquad would have been a zipper-noise generator on its own, so the filter core changed too. A direct-form biquad whose coefficients jump every sample is not a well-defined time-varying filter - its internal state means something different after each jump. Every band (bell, low shelf, high shelf) instead runs on lnct::TptSvf, a topology-preserving-transform (trapezoidal- integration) state-variable filter built from Andrew Simper's published Cytomic equations, with no third-party code vendored. Its two state variables are physical integrator outputs that keep their meaning no matter how the coefficients move, so gain can be modulated every sample without artefact. The bell's frequency warp is completely gain-independent - dynamic gain enters only as a scalar bandpass mix, never by re-tuning the centre frequency - and integrator state is kept in double even though the audio path is float, because the trapezoidal update's own cancellation loses digits in single precision at low frequency and high Q.

The realised static response is unchanged by any of this: verified against an independent double-precision reference to better than -100 dBFS peak residual over ten seconds of broadband noise per setting, and against the analytic RBJ magnitude response to within ±0.05 dB across a band-type/gain/Q grid. At exactly 0 dB every gain-dependent mix scalar is exactly zero, so an idle band is bit-transparent by construction, with no branch and no special case.

Detector design

In Split mode, the detector cascades two bandpass biquads at the band's own frequency/Q rather than one, because a single stage only reaches about -12 dB two octaves out at Q 1 - not enough headroom against a loud out-of-band tone falsely triggering the band. The cascade measures better than -24 dB, comfortably past the plugin's own >20 dB two-octave isolation bar. Stereo (or wider) input is linked: each cascade stage runs independently per channel with its own filter state, but the envelope follower is a single band-wide value fed by the loudest (max-abs) sample across channels at every instant - which avoids the stereo-image shift that fully independent per-channel gain reduction would introduce. There is no per-band unlink option. The envelope itself is a one-pole peak follower run per sample for correct ballistics timing; only the bandpass's own coefficients are throttled to sub-block granularity, never the envelope.

Auto-release: a dedicated fast reference envelope

Auto Release (off by default) shortens a band's effective release whenever the signal's own envelope is already falling on its own, clamped so the result is always at least as fast as - never slower than - the manual Release setting. The detail worth knowing: the fall-rate measurement comes from a second, dedicated, always-fast envelope inside the detector (same Attack coefficient, but a fixed release tied to the plugin's own 5 ms floor), not from the main envelope. Deriving it from the main envelope does not work - a slow envelope is itself a low-passed view of the input, rate-limited to roughly its own time constant, so "how fast is the slow envelope falling" mostly measures the slow envelope's own coefficient back at itself. An earlier internal implementation attempt made exactly that mistake, and its output was measurably identical whether Auto Release was on or off.

Anti-aliased saturation, still at zero latency

The optional per-band Saturation stage replaces a plain tanh waveshaper with a first-order antiderivative-antialiased (ADAA1) kernel of the same shape, computed as the difference quotient of ln(cosh(x)) between consecutive samples - same harmonic character, measurably less fold-back, no oversampling stage and no added latency. The specification is deliberately relative rather than an absolute alias-floor claim; see "Latency and aliasing" below for the measured numbers and the reasoning. Saturation only ever engages while a band is actively boosting (static + dynamic gain net positive) - a cutting or idle band is bit-identical with it switched on.

Engineering hygiene

• Zero heap allocations on the audio thread, proven under a replaced allocator, including with the sidechain bus active and Auto Release / Gain-Q / Saturation engaged on every band. Coefficient updates write a stack-only ArrayCoefficients result directly into already-allocated storage instead of calling juce::dsp::IIR::Coefficients::makeXxx(), which heap-allocates a fresh object on every call.
• Idle work is skipped. The detector bandpass and the SVF frequency warp recompute only when the smoothed frequency/Q have actually moved past an epsilon, and the per-sample SVF scalars are memoised on an exact gain match, so a band sitting at a settled gain recomputes nothing and is bit-transparent once its parameters are static.
• NaN/Inf recovery is designed in, including on the sidechain. A non-finite value would otherwise poison the SVF's integrator recursion permanently; the filter snaps its state clean and returns 0, recovering within the same block rather than needing a reset() - verified with a NaN/Inf-poisoned sidechain and all six bands set to External, which cannot silence or poison the main output.

Parameter reference

Per band (Band 1 - Band 6, identical controls unless noted)

Per-band voicing defaults (docs/voicing-notes.md) - Q/Attack/Release tuned to each band's typical role along the existing frequency ladder (v0.3.0), Threshold calibrated by measurement so that every band begins engaging at the same programme loudness (issue #4 calibration pass - see below):

Each band's default Threshold equals that band's own measured detector level under a -18 dBFS RMS pink-noise programme anchor (the common digital alignment-level convention), so at typical mix level every band sits right at the edge of engaging: dial in a Range and the band immediately starts working on anything at or above normal programme loudness, on every band alike. Run material hotter or colder than that convention (or into an Input Trim adjustment) and the engagement point shifts with it - Threshold remains an ordinary, per-band control; only its default is calibrated.

Global

Presets

A preset bar sits at the top of the editor: [<] [Preset Name] [>] to step through the factory and user library alphabetically, Save/Save As... to write your own, Delete for user presets, Import.../Export... for single .basilicapreset files or .zip banks, and a menu (click the preset name) with a "Set current as default" entry for your own out-of-the-box starting point. Eleven factory presets ship - see docs/presets.md for what each one does and why. User presets are stored per-user at ~/Library/Audio/Presets/Yves Vogl/Lancet/ on macOS (%APPDATA%/Yves Vogl/Lancet/Presets/ on Windows).

The editor's frame strings (preset bar labels, menus, dialogs) are localised to German automatically when the system language is German; parameter names, units, and technical terms (Attack, Release, Hz, dB, ms, …) always stay in English, matching every other Basilica Audio plugin.

Tips

• Set Gain and Range separately, deliberately. Gain is what the band does always; Range is what it does in addition, only when triggered. A band with Gain=0, Range=-6 is silent at rest and only cuts when the resonance flares - very different from Gain=-3, Range=-3, which is always cutting a bit and cuts harder when triggered.
• Use Listen to find the problem before you set Threshold. Sweep Freq/Q with Listen engaged until you clearly hear the resonance or harshness in isolation, then set Threshold to just above where it sits when it's not a problem - this is far more reliable than guessing at a Threshold value against the full mix.
• Narrow, high-Q bands with negative Range are the classic resonance-taming setup (a boxy 300-500 Hz build-up, a harsh 2-4 kHz pick/reed edge, a sibilant 6-8 kHz de-esser band). Keep Q high enough that the cut doesn't audibly thin the surrounding tone when it engages.
• Wide, low-Q bands with negative Range make a gentler, broader dynamic tone control - useful on a mix bus for taming a whole register (e.g. "the low-mids get a bit much whenever the whole band hits together") without the surgical narrowness of a de-esser-style band.
• Positive Range (upward/duck-in) is the less obvious move - try it on a low-Q high-frequency band to bring out pick attack or breath/consonant detail only on the notes that need it, rather than boosting the whole register (and its noise floor) all the time.
• Fast Attack + fast Release can pump audibly on sustained material (bass, pads, sustained vocal notes) - if a band sounds unstable or "breathing," try a slower Release first before reaching for a narrower Q.
• Mix below 100% keeps the dynamic move's character while reducing its depth - a quick way to dial back an over-aggressive Range setting without re-tuning every band's Threshold/Range from scratch.
• Reach for SC Mode: Wide when you want a band to breathe with the mix, not police one resonance (v0.4.0). Split is the right default for surgical work - only content near Freq can trigger the band. Wide makes the band respond to overall level while still only moving its own frequency region, which is what you want for "duck the low-mids whenever the whole band hits" or for a band tracking a full-range sidechain source. A Wide band with a narrow Q is a perfectly sensible combination: the Q shapes what moves, the mode decides what triggers it.
• With an external sidechain, set Threshold against the sidechain's level, not the main signal's (v0.4.0). Once SC Source is External the band is reacting to whatever your host routes in, so the level reaching Threshold has nothing to do with what Lancet is inserted on. Engage that band's Listen to hear the sidechain feed directly while you find the Threshold - Listen always auditions whatever the detector is actually hearing, including the sidechain.
• Time-align the sidechain in your host if it matters. Lancet inserts no alignment delay anywhere (see "Latency and aliasing" below), so a sidechain that arrives late in your host arrives late at the detector too.
• Now that Attack is true, try it before you reach for anything else (v0.4.0). If a band grabs too hard on transients, a slower Attack is a real control again rather than a no-op below ~50 ms - that is usually a better first move than widening Q or backing off Range.
• Saturation is only worth switching on where a band boosts. It is scoped to boosting bands by design, so enabling it on a cutting de-esser band does nothing at all. Pair it with a modest positive Gain (and optionally a positive Range) on a body/warmth band.

Latency and aliasing

Lancet adds zero latency - always. It reports 0 samples before and after your host prepares it, at every sample rate and block size, with every band engaged and with the sidechain bus enabled. That is verified by impulse response (the peak really comes back on sample 0), not merely reported, and there is no dry-path delay compensation anywhere in the plugin. Every filter in the signal path - the six bands and their detectors - is minimum-phase with no lookahead.

Two v0.4.0 decisions keep it that way, deliberately:

• The Saturation stage is anti-aliased arithmetically rather than by oversampling. It uses a first-order antiderivative-antialiased waveshaper of the same tanh shape the previous versions used, so it has the same harmonic character with measurably less fold-back - and no oversampling stage, which would have cost either latency or phase distortion. Measured suppression of the dominant 30 kHz to 18 kHz fold at 48 kHz is 13.2 dB at -24 dBFS, 10.0 dB at -12 dBFS and 8.4 dB at -6 dBFS input. No absolute alias floor is claimed, and that is honest rather than coy: first-order treatment at the base sample rate does not earn one. If you drive a band hard on very high-frequency content and want a guaranteed absolute floor, that is what an oversampling plugin is for, and oversampling costs latency.
• The external sidechain carries no alignment delay, so it must already be time-aligned by the host - the same requirement comparable dynamic EQs have.

Sample rates and formats. AU, VST3 and Standalone. Verified finite and zero-latency at 44.1, 48, 88.2, 96, 176.4 and 192 kHz, and across a sample-rate change mid-session. Mono and stereo main layouts are both supported; the sidechain input can be disabled, mono or stereo independently of the main layout, and anything wider than stereo on it is rejected rather than silently misinterpreted.

Sessions, presets and compatibility

Sessions saved by any earlier version load cleanly, with every stored parameter value preserved exactly and every parameter added since then sitting at the default that reproduces the older behaviour. Concretely: a v0.1.0 session's values survive, the v0.2.0/v0.3.0 per-band toggles (Auto Release, Gain/Q, Saturation) come back off, and v0.4.0's twelve new per-band choices come back at Internal and Split - the only routing any previous version ever had.

Saved state carries a schema-version stamp from v0.4.0 onward. A state without one is read as the older schema. Nothing needs converting today (every newer parameter's default is the older behaviour), but the stamp means a future release that genuinely does need to convert something has a reliable way to tell what it is reading.

One caveat, and it is the important one: while your stored values are untouched, a session that used a non-zero Range together with a fast Attack will sound different under v0.4.0, because the Attack path was broken before and is now fixed. See "What's new in v0.4.0" at the top of this document. The same applies to the ten factory presets that predate v0.4.0 - none of their stored values changed, but the ones whose names promise speed (De-Ess Stack, Transient Snare Crack, Fast-Recovery Demo) now behave the way their names always claimed.

Known limitations

Stated plainly, because knowing them is more useful than not:

• The GUI is a functional slider/toggle/combo editor. The custom vector-drawn interface with per-band gain-reduction needles is a later milestone. The plugin already measures what those needles will show, and that measurement is verified against the gain reduction actually present in the output - but nothing displays it yet.
• Auto Release, Gain/Q and Saturation have no editor controls. All three are fully automatable and fully preset-controllable; dedicated toggles arrive with the GUI pass. SC Source and SC Mode are the deliberate exceptions, because a sidechain routing you cannot select from the editor is not usable at all.
• Q is ignored in Shelf mode. Band 1's Low Shelf and Band 6's High Shelf always use the standard flat shelf slope (Q 0.707), for both the audible filter and its matched detector bandpass.
• No absolute aliasing floor is specified for the Saturation stage - see "Latency and aliasing" above.
• Detection is stereo-linked with no unlink, and there is no per-band M/S or L/R placement - see "Detector design" under "Under the hood" above.
• There is no ratio control, by design. Above the knee the gain moves 1:1 with how far the detector has overshot Threshold, until it reaches Range - which acts as a hard ceiling on the dynamic depth. Range is the "how far can this move" control; the knee (whose width scales with Range) provides the ramp-in.
• The per-band voicing defaults are engineering judgment, not ear-tuned against reference material. The direction of the v0.3.0 voicing table (low frequency slow and gentle, high frequency fast and surgical) is an established mixing convention, and the resulting ballistics ordering is measured and frozen by the test suite. The exact numbers, and the Saturation drive curve, are tuned judgment - not validated against real vocal/guitar/mix material and not calibrated against any other product. A by-ear pass against reference-class dynamic EQs on real programme material is a named, still-open item; see docs/voicing-notes.md for the full honesty section on exactly which numbers are which.
• Not yet in the plugin, tracked for later releases: opt-in spectral resonance suppression, per-band M/S and L/R placement with stereo unlink, optional HF de-cramping (deliberately deferred, because de-cramping changes the static response of existing sessions), wider Range/Q/Release ranges (which need their own migration story, since remapping a parameter range changes host automation-curve mapping), lookahead, per-band ratio, linear phase, more than six bands, and a spectrum-analyzer/EQ-curve display.
• Lancet is pre-1.0. Release binaries for macOS are Developer-ID-signed, notarised and stapled; Windows binaries are not yet Authenticode-signed. Breaking changes are possible until v1.0.0; installers are a later milestone. See README.md.