Skip to content

Implementation plan — Harmony Formant via per-voice biquad (Option B)

Status: experiment reverted (2026-08-03) — the per-voice biquad formant path (option B) was implemented and A/B-tested, but its peaking-EQ formant shift was too subtle to be an audible Harmony Formant control and it clicked at position 0. User decision: keep the granular method (HC.13) as the sole Harmony Formant path; the biquad option was removed from the source. This plan is retained as a record of the experiment.

Goal

Replace the harmony voices' granular formant preservation (grain read-speed scaling inside PitchShifter) with a filter-based formant shift (per-voice ovtdsp::FormantPreserver biquad), so the Harmony Formant knob works fully and stably on all harmony types — including single octave-shifted voices (Drone, Octave Below, Octave Above), which are the ones that artifact with the granular method at extreme values (−5).

Context / why

  • Harmony formants are currently shifted inside PitchShifter::process by scaling the grain read speed: grains[v].speed = F, F = 2^(harmonyFormant/12) (Source/dsp/PitchShifter.cpp).
  • On a voice that is also strongly pitch-shifted (octave → pitch ratio 2.0 / 0.5), the granular read-speed formant breaks the OLA COLA sum → wobble / rapid pops, lateralized per ear (HC.13).
  • Option A (blend the granular formant ratio toward 1.0 by pitch deviation) reduced the artifact but cannot deliver a full −5 formant shift on octave voices — it reduces exactly where the user wants it. It also made the knob effectively dead on mono octave types (regression, mitigated by a 0.5 floor).
  • Option B applies the formant shift with a stable biquad filter after/before the pitch shift, decoupled from the granular read-speed, so it works uniformly regardless of the voice's pitch ratio.

Proposed design — B1: pre-warp biquad per voice (matches the lead voice)

The lead voice already uses the proven pre-warp path (Source/PluginProcessor.cpp): 

formantPreserver.setFormantShift(shiftSemitones);
formantPreserver.process(buffer, ratio);       // pre-warp input formants
pitchShifter->process(buffer, ratio, userFormantRatio, f0);
We mirror this per harmony voice, with the Harmony Formant value as the shift and the granular formant ratio forced to 1.0.

Per shifted voice v: 1. Copy the shared input synthWorkBuffer into a per-voice staging buffer. 2. Pre-warp: formantPreserverHarmonyVoices[v].process(staging[v], ratioH) with setFormantShift(harmonyShiftSemitones). 3. Pitch shift with no granular formant: shiftedVoicePitchShifters[v]->process(staging[v], tmp, ratioH, 1.0f, safe_f0).

Net effect: the voice's pitch is shifted by ratioH and its formants end up at the harmonyShiftSemitones offset — identical intent to the current granular method, but using the stable biquad, and it works at octave shifts.

Design decision recorded

  • B1 (pre-warp, above) is preferred because it preserves formants relative to the source (an octave-up voice keeps human formants, just offset by the knob), matching the lead voice and the current granular behaviour.
  • B2 (post-shift): applying the biquad to the already-pitch-shifted tmp would shift the chipmunk formants (shifted by the pitch ratio) rather than preserve them → different/incorrect sound for octave voices. Rejected as primary, kept as a fallback only if B1 sounds wrong.

File changes

Source/PluginProcessor.h

  • Replace the single formantPreserverHarmony member with: std::array<ovtdsp::FormantPreserver, maxShiftedVoices> formantPreserverHarmonyVoices; (each voice needs its own filter state — a shared instance would bleed state across voices).
  • Add per-voice staging buffers for the pre-warped input: std::array<juce::AudioBuffer<float>, maxShiftedVoices> shiftedVoiceFormantInput; (resized in the voice loop like shiftedVoiceBuffers).

Source/PluginProcessor.cpp

  • prepareToPlay(): loop over formantPreserverHarmonyVoices, call prepare(sampleRate, samplesPerBlock) and apply the same mode/strategy/ voiceType/Q/smoothing configuration currently applied to formantPreserverHarmony (lines ~2324–2387).
  • reset(): call reset() on each voice instance.
  • syncParameters() / the formant config block: configure all voice instances instead of the single one.
  • Voice loop (Source/PluginProcessor.cpp ~line 2608):
  • resize shiftedVoiceFormantInput[v], copyFrom(synthWorkBuffer).
  • formantPreserverHarmonyVoices[v].setEnabled(true), setFormantShift(harmonyShiftSemitones), process(shiftedVoiceFormantInput[v], ratioH).
  • call shiftedVoicePitchShifters[v]->process(shiftedVoiceFormantInput[v], tmp, ratioH, 1.0f, safe_f0).
  • Revert Option A: remove the voiceFormantRatio blend (HC.13); always pass granular formant ratio 1.0f. Option A is superseded by Option B.

Configuration reuse

The existing formant config block (mode / strategy / voice type / Q multiplier / smoothing alpha) already targets the harmony path. It is re-targeted from the single formantPreserverHarmony to the per-voice array. The harmony_formant APVTS parameter value (harmonyShiftSemitones) is unchanged and feeds setFormantShift.

CPU / performance

  • Adds up to 4 biquad pre-warp passes + 4 buffer copies per block (one per active shifted voice). The lead already does 1 pass; 4× that is modest for modern CPUs. Mitigation if needed: reuse the per-voice tmp buffer for staging in-place is not possible (input must stay pristine for other voices), so staging buffers are required.

Risks & validation

  • Formant character may change vs the granular method on close harmonies. Validate by ear across all types at multiple Harmony Formant values (±5).
  • Octave voices (Drone / Octave Below / Octave Above): confirm the wobble/pop is gone and the knob is clearly audible.
  • Formant direction: verify the knob moves formants in the expected direction (down at −5) on both octave-up and octave-down voices.
  • CPU: check the standalone doesn't hit the real-time budget.
  • If B1 produces an unexpected sound, try B2 (post-shift on tmp, passing ratio = 1.0f to process to skip pitch compensation).

Rollback

Option A's blend is small and localized (one expression + comment). If Option B is unacceptable, it is easy to restore by keeping the granular formant ratio path and removing the per-voice biquad. Both paths can coexist behind the formant_strategy parameter if desired, but the default should be a single stable path.

Comparison option (A/B at runtime)

To validate B without losing the current sound, add an APVTS AudioParameterChoice harmony_formant_method:

  • "Granular (current)" — the existing HC.13 path (per-voice formant-ratio blend, floor 0.5) via the grain read-speed in PitchShifter.
  • "Biquad (per-voice)" — the new HC.14 per-voice FormantPreserver pre-warp.

The voice loop branches once per block on the selected method. This lets the user flip between the two while singing (no recompile / reload), which is the cleanest way to validate that the biquad path both removes the wobble/pop and keeps the formant knob fully working on all types (including Drone / Octave Below / Octave Above).