Seraph

A doubler that holds its interval. Micro is a dual-head constant-ratio delay-line shifter: a delay whose length ramps at a fixed rate produces exactly the target pitch ratio, read with cubic Catmull-Rom interpolation and crossfaded so whichever head is wrapping is silent at that instant. Measured accurate to 0.5 cents at ±30 cents — and to a −100 dBFS null against a plain static delay at zero detune, because at zero detune that is exactly what it becomes. The sweep runs upward from the base delay rather than centred on it, because centring would mean reading into the future; the consequence — voices averaging 34–49 ms, slappier than the Classic engine — is documented, not hidden.

A spectral shifter configured against a budget, not against its own defaults. Shift mode runs one instance of a vendored, MIT-licensed phase vocoder per voice, configured with a 30 ms window at a 7.5 ms hop — specified in seconds, never as a bin count, so a 96 kHz session keeps the same physical window instead of silently halving it. The engine's own default preset would use roughly 150 ms, which is out of budget for a tracking-vocal insert. It is vendored rather than hand-rolled for a reason worth stating: LPC envelope estimation is at its worst on choir-register vocals, and TD-PSOLA needs a reliable pitch mark per period — both assume a monophonic source, and this plugin is routinely fed stacks and buses.

Humanisation that is decorrelated, not unrepeatable. Three slow random walks per voice — timing ±10 ms, pitch ±3 cents, level ±1.5 dB — from a seeded generator through a 0.5 Hz one-pole, on a fixed control clock that never consults the host's block size. Two renders from the same reset state are bit-identical, and a host handing over 64 samples at a time produces exactly the same drift as one handing over 256. At 0% every offset is exactly zero, so the Classic engine stays bit-identical to the previous release.

Lookahead de-essing that attenuates instead of boosting. The de-esser works by adding a scaled copy of the detected band back onto the signal, which only reduces if the two are time-aligned. Delaying the audio and leaving the band undelayed — the obvious implementation — would misalign them by the lookahead length, and because sibilance is decorrelated at a 2 ms lag the "subtraction" would add roughly 0.8× the band's power back in. Seraph delays both, runs the detector on the undelayed band, and pushes the gain through a monotonic sliding minimum so it reaches its target before the ess arrives. The onset overshoot a 1 ms attack otherwise lets through is gone: ≤ 0.5 dB of excess over the settled level, versus over 3 dB without.

Latency is zero only in the default configuration. Classic and Micro report 0 samples, and that default is unchanged from the previous release. Shift mode reports 1440 samples (30.0 ms at 48 kHz, scaling with sample rate), and De-Ess Lookahead adds round(ms × sample rate / 1000) samples on top:

A click pushed through the plugin arrives where the reported number says it will to within one sample, in every mode — the doubler's main path, the de-esser's band, and the dry capture feeding the Mix crossfade are each delayed to match. Both parameters that can change the number are non-automatable, so a host only ever sees the change as a deliberate move, masked by a 10 ms fade.

Compatibility as a merge gate. A pre-existing session renders byte-identical after upgrading, at 44.1 / 48 / 96 kHz and block sizes 64 and 1024. All eight new parameters default to the value that reproduces the previous release exactly; state carries a schema version; the new parameters are appended after the frozen existing set so index-addressed automation lanes cannot silently re-point; no parameter ID was renamed or removed; and every existing factory preset loads with the new keys exactly on their defaults.

Honest engineering, enforced by CI. 108 test cases on macOS and Windows: zero heap allocations on the audio thread under a replaced allocator — including live mode switches inside the guard — a −90 dBFS neutral null across six sample rates from 44.1 to 192 kHz, Shift mode processing a second of stereo in under 100 ms, and pluginval at strictness 10 plus auval -strict on every push.