Basilica Audio

Miserere — user manual (v0.7.0)

Four voices, one prayer — the parallel vocal template in a single unit.

What Miserere is

Miserere packages the documented 2010–2023-era parallel vocal template — the "rough vocal" workflow popularized in public interviews by mixers such as Andrew Scheps — into one plugin: a Direct path plus four parallel return busses (CRUSH, SANDWICH, SPREAD, SLAP), each with its own return fader, Mute and Audition. This is a documented, publicly sourced technique from that era (see research-notes.md), not an endorsement by or association with any named person or brand.

The core idea, and v2's correction over v0.1.0: the Direct path is a wire. Out of the box, every optional section on it is OFF, so the dry vocal passes through essentially untouched — its natural envelope and phrasing survive. Everything else is layered underneath it via the four return busses, which are copies of the direct-path output at unity, processed hard, and blended back in at a modest level. "Even with all that stuff in the mix, you'd probably think the vocal is bone dry" is the calibration target.

Signal flow

in → [In Trim] → DIRECT PATH (serial; every section optional, ALL OFF by default:
                   De-Esser (pre) → FET Comp light → Console EQ → Sat → De-Esser (post))
        │ = "the channel". Output feeds the sum at unity AND all four sends (unity taps):
        ├─→ ① CRUSH    : FET limiter, all-buttons character        → return fader
        ├─→ ② SANDWICH : Passive EQ → Opto Leveler → Passive EQ    → return fader
        ├─→ ③ SPREAD   : dual micro-pitch (≈30/50 ms, ±cents, L/R) → return fader
        └─→ ④ SLAP     : ≈110 ms dark single-repeat delay          → return fader
   Σ (direct + returns) → [Parallel macro trim scales returns ①–④] → [Out Trim]
                        → [Output Limiter — off by default] → out

Busses ①/② are minimum-phase and add zero latency, so they stay sample-aligned with the direct path — parallel summing never combs regardless of settings. Busses ③/④ are delays by design (see architecture.md). See research-notes.md for the sourced findings behind every default below.

The Direct path

Off by default, section by section, in signal order:

Since v0.5.0 the Drive is a transformer model rather than an added harmonic term: the signal runs through a flux integrator into a biased saturator and back out through that integrator's exact inverse. Because magnetic flux scales as voltage over frequency, the third harmonic rises toward the bass on its own (measured +12 dB going from 100 Hz to 50 Hz) instead of being weighted by hand. At 0 dB the Drive is a bit-exact bypass. The 12 kHz shelf is also magnitude-matched since v0.5.0, so at 44.1 kHz it keeps its analog shape in the top octave instead of being squeezed toward Nyquist.

The four return busses

① CRUSH — FET limiter, all-buttons character

No threshold knob: Input drives the signal into a fixed per-ratio threshold and knee. Ratio selects 4:1/8:1/12:1/20:1/ALL (ALL is a plateau-shaped curve with a deliberate give-back and a short attack lag that lets transients punch through before clamping — the "snap"). Attack/Release are 1–7 dials where a HIGHER number is FASTER, matching the hardware convention this is modelled on; release is program-dependent (fast after brief transients, several times slower after sustained heavy compression). Style selects the limiter's voicing: All-Buttons (default), a softer, fixed 2:1 Gentle voicing, or — since v0.6.0 — Vintage, an early-revision hot-bias state that keeps the full ratio row active but bites two dB earlier, runs the feedback loop hotter and more than doubles the FET cell's residual second-harmonic "hair", for a rattier, more coloured crush at the same settings. This bus is meant to sound "terrible" soloed (use Audition) and good blended in.

CRUSH also carries a touch of program-dependent colour: as gain reduction builds, a transformer-style low-frequency saturation and the FET cell's own residual second harmonic blend in on top of the limiter's detector-ripple character — negligible at light settings and growing only as the bus works harder. A clean, barely-compressed signal is unaffected; lean on Input and Audition to hear it come alive.

Since v0.5.0 none of the above is tabulated. The detector is a real feedback loop: the sidechain is driven from the bus output through a single-capacitor RC network with Attack in the charge path and Release in the discharge path, which is why the behaviours this style of limiter is known for now simply happen rather than being scripted — the effective ratio rises as a note is held, the Release setting audibly changes how fast the attack arrives, the knee tightens as you go up the ratio row, and ALL overshoots before settling. ALL is its own setting, not an interpolation between the numbered ratios.

② SANDWICH — Passive EQ → Opto Leveler → Passive EQ

Two independent Passive EQ instances bracket an opto-style leveler. Each Passive EQ offers a shared-frequency LF Boost and Cut (both can run at once — a deliberately non-cancelling curve, not a simple sum to flat), an HF Bell Boost with variable bandwidth, and an HF Shelf Atten. The Opto Leveler has no threshold: Peak Reduction drives the cell harder, with Limit tightening the compression toward limiting. Emphasis makes the detector progressively HF-selective (up to −10 dB less LF sensitivity), so at high settings it reacts mostly to sibilance/presence, "like a multiband". Residual (default on) keeps the Passive EQ's small, never-fully-flat vintage tilt; defeat it for a cleaner EQ.

Since v0.5.0 the leveler is a photocell model rather than a set of drawn curves. The compression ratio, the two-stage release (a quick initial recovery followed by a long tail), the memory effect — hold it down longer or harder and it lets go more slowly — and the program-dependent attack are all consequences of how charge carriers in the cell build up and drain away. There is no ratio control because there is no ratio parameter in the circuit; what you hear is the cell's own behaviour, which is the whole appeal of this style of leveler.

Since v0.6.0 a Colour switch selects the cell's era: Classic (the default — the photocell calibration exactly as before), Quick (a later solid-state-era optical voicing: faster recovery, much less release memory, an almost clean output stage) and Deep (an earlier-era voicing: slower recovery, a longer memory tail, and a thicker tube/transformer output stage). The switch changes the cell's charge-carrier kinetics, not its level calibration — all three land in the same gain-reduction class at the same Peak Reduction setting, and swapping colours mid-performance is click-free.

The LF Boost and Cut network is likewise now the hardware ladder's exact response. Two practical consequences: running Boost and Cut together gives the classic low-end shape (a lift underneath with a dip just above it) because the cut corner genuinely sits above the boost corner, and the Cut on its own is broad — at full attenuation on the 100 Hz setting it is still about 1.6 dB down at 2 kHz. That breadth is the circuit's, not a bug; it is why the control is normally used against the Boost rather than alone.

③ SPREAD — dual micro-pitch

Two short delay taps (~30 ms pitched up, ~50 ms pitched down), hard-panned L/R. Detune sets the pitch offset in cents (default 6 — deliberately small, so the ear reads "pushed to the outside" rather than chorusing). Time scales both base delays together; Width blends from a fully centred sum (0%) to the full hard pan (100%).

Since v0.5.0 both shifter delay lines are read with third-order Lagrange interpolation instead of a linear read, which recovers 0.90 dB of high-frequency content at 10 kHz that the linear read was losing; the grain crossfade is longer and equal-power, which measurably lowers the periodic level ripple on sustained tones; and Detune and Time are smoothed per sample, so automating either no longer steps at block boundaries.

Since v0.6.0 the splice is period-adaptive: a pitch tracker (autocorrelation on a lowpassed, decimated copy of the input — a few percent of the bus's CPU) watches the material, and when it is confidently pitched, the spacing between each shifter's two crossfading taps snaps to a multiple of the note's period while the crossfade law shifts to amplitude-complementary. Both taps then reinforce at every harmonic instead of interfering, which removes the note-dependent level ripple sustained tones used to meet (see Known limitations). Unpitched material leaves the tracker below its confidence gate and passes through the identical v0.5.0 path. The tracker only ever reads signal history — reported latency stays 0.

Also since v0.6.0 the crossfade geometry stays causal at short Time settings: below 100% the tap spacing shrinks with the base delay (down to ~15 ms at 50%), which slightly quickens the splice cadence. Previously a tap could sit against the delay line's lower bound for stretches of a second or more and leak an unshifted copy of the input into the return.

④ SLAP — single-repeat dark delay

Time (50–160 ms, default 110 ms, plain milliseconds — deliberately not tempo-synced). Feedback is fixed at 0 in v2: there is exactly one repeat, and its darkness comes from a built-in tape-style voicing (Tone sweeps a progressive HF loss plus soft saturation baked into that single repeat) rather than a filtered feedback loop. Stereo switches from the default mono return (the classic mono slap behind a stereo-widened vocal) to independent L/R delays.

Since v0.5.0 the repeat is voiced as an actual tape transport rather than a filtered delay: the saturation now sits on the record side (ahead of the delay write) instead of being lumped onto the tap, which is why repeats and input no longer brighten together, and a fixed head bump adds a small low-mid lift.

Wobble (default 0%) is the transport's wow and flutter — a slow pinch-roller waver, a faster capstan flutter, and a slow random drift, each wandering independently so it never settles into an obviously repeating pattern. The dial spans roughly 0 to 0.5% wow-and-flutter; small amounts (10–25%) read as "this was on tape" without sounding broken, and high settings get seasick on purpose. At 0 the modulation is genuinely switched off, not merely turned down.

Age (default 0%) is tape wear: hiss with an asperity component that rides the signal (the noise breathing with the vocal is most of what makes it read as tape rather than as added hiss), plus extra head-to-tape spacing loss that dulls the repeat further as the dial rises. It affects only the SLAP return, never the direct path. At 0 nothing is generated at all.

Fader logic

External sidechain (v0.7.0)

Miserere exposes an optional Sidechain input bus, disabled by default — enable it in your host's plugin routing, then switch individual detectors over to it with the Ext Key lamp in the Direct Path, Crush and Sandwich panels. There is no master switch: each detector decides for itself, so you can key CRUSH from a snare while SANDWICH keeps listening to the vocal. If the bus is absent or disabled, or the host sends no key, every switch falls back to internal detection silently. A mono key feeding a stereo instance keys both channels.

What keying actually does to CRUSH and SANDWICH — read this before assuming it is "the same sound, different detector source". Both of those busses are feedback designs, and that is not an implementation detail, it is where their character comes from:

An external key does not add a detector input to those loops. It replaces the loop drive, which converts both modules into feed-forward keyed compressors for as long as Ext Key is engaged. The photocell physics, the ballistics and the colour stages are unchanged, but the static curve is not: a feedback detector sees the already-reduced signal and backs itself off, while a feed-forward detector sees the full-scale key and does not. Measured on the same signal at the same settings (a −22 dBFS tone, 0 dB Input), CRUSH produces about 3 dB of gain reduction with internal detection and about 21 dB when keyed with a copy of its own input. That is a legitimate and useful mode — it is simply a different compressor, and expecting your internal-detection settings to carry over unchanged will surprise you.

The Direct FET is the exception: it has always been feed-forward (its envelope comes from the pre-gain input), so keying it really is a pure detector-source swap — same topology, same curve, different source.

The key is read only. It never reaches the audio path, it is not part of the sum, and it does not affect the zero-latency guarantee.

Output limiter (v0.7.0)

The last stage in the plugin, after Out Trim, with its own needle meter on the Global panel. Off by default, and while it is off it is a bit-exact bypass — the default wire stays a wire.

This is a sample-peak ceiling, not a true-peak ceiling — and it deliberately cannot be one. True-peak limiting means detecting the reconstructed waveform between the samples, which requires oversampled detection, whose filters are delays; honouring their verdict means holding the audio back by that delay, i.e. lookahead. The zero-latency guarantee (see below) rules that out. What the limiter guarantees is that no output sample exceeds the Ceiling; the analogue waveform your converter reconstructs between those samples still can.

Measured with an 8× windowed-sinc reconstruction (tests/OutputLimiterTests.cpp, where these figures are regression-frozen), at a −0.3 dB ceiling:

Programme Inter-sample overshoot above the ceiling
11.025 kHz tone at 44.1 kHz, phased so every sample straddles a crest (the analytic worst case for a sine) 3.01 dB
11 kHz tone, arbitrary phase 0.71 dB
1 kHz tone, 5 ms release 0.06 dB

Practically: on real programme material the overshoot is a fraction of a dB, and the 3 dB figure is the mathematical ceiling of what a sine can hide between samples, not a typical result. If you are delivering to a specification that mandates a true-peak limit (−1 dBTP for lossy encoding, for instance), set Ceiling with that headroom — or use a true-peak limiter at the end of your master chain, where its latency costs nothing.

Presets

A preset bar sits at the top of the editor: [<] [PresetName*] [>] [Save] [Save As...] [Delete] [Import...] [Export...]. Clicking the preset name opens a Factory/User menu; a trailing * means the current preset has unsaved changes. Thirteen factory presets ship in the box (see presets.md for what each one is for) — including Tape Slap 7.5 and Worn Slap, added in v0.5.0 to exercise Wobble and Age, and BV Mode, a background/stacked-vocal starting point with every return pushed harder than the lead template; user presets save to ~/Library/Audio/Presets/Yves Vogl/Miserere/ on macOS (%APPDATA%/Yves Vogl/Miserere/Presets/ on Windows). The preset menu's "Set current as default" makes any preset — factory or user — load automatically on every fresh instance; "Import..." accepts both single preset files and zip preset banks.

Starter recipe

  1. Leave the Direct path off, or add De-Ess Pre / a touch of Console EQ if the source needs it. Keep FET Comp and Sat off unless the vocal specifically needs light insert compression.
  2. CRUSH starts at −9 dB by default with the ALL-Buttons character already engaged — bring Input up until Audition shows heavy, "disaster in solo" compression, then trust the default fader level and adjust by ear from there.
  3. SANDWICH starts at −12 dB; raise Peak Reduction until the vocal thickens without audibly pumping in context.
  4. SPREAD and SLAP (−18 dB / −15 dB by default) should each pass the "you only notice it's gone when you mute it" test — if either is audible as a discrete effect, pull it back.
  5. Use Parallel to back the whole layer off quickly on quieter/more organic material.

Latency and aliasing (v0.5.0)

Miserere reports and adds 0 samples of latency, at every sample rate and with every section engaged. That is a deliberate constraint, not an oversight: the four return busses are summed against a bit-transparent direct path, so anything that delayed one bus by even a fraction of a sample would comb the sum.

Keeping that promise rules out oversampling, which is how most plugins tame the aliasing that saturation produces. Instead every drive stage here computes its distortion in a form that suppresses aliasing arithmetically, split so that the clean part of the signal passes through exactly aligned and untouched — no delay, and no high-frequency dulling at 44.1 kHz.

What that buys, measured at 44.1 kHz: on a programme-realistic probe (a 3 kHz tone at −12 dBFS through hot-but-musical drive settings) non-harmonic content sits at or below −60 dBFS. Pushed to deliberately unrealistic extremes — a full-scale 12 kHz tone into maximum drive — the treatment still removes at least 12 dB of aliasing versus the untreated curve, but no absolute floor is claimed there, and you should not expect one from any zero-latency design. If you want a saturated 12 kHz sine at full scale to stay clean, that is what oversampling is for, and it costs latency.

Known limitations