Conversation
Delays the wet signal before it reaches the convolver, so an IR's tail starts later without touching the loaded impulse response itself (padding the IR would force a full convolver rebuild with its silent install fade on every live change, and inflate the uniform engine's kernel-length-bound CPU cost and its short/long classification). Always runs at kChainBaseSampleRate, inside the block's base-rate island, right before Convolution::process() - the same rate convolution itself always runs at regardless of oversampling. A preallocated ring buffer (sized once for the 1s max in prepare(), zero allocation after) is read with linear interpolation against a smoothed delay-in-samples target, so live changes glide instead of stepping. The live-change ramp (10ms) was tuned by ear. Range 0-1000ms, wired through setBlockParam (normalized 0..1, same convention as inputGain/outputGain/mix) and persisted the same way. Predelay pushes an IR's audible tail out in time, so the reported host tail length now accounts for it too (refreshIrTailLength). Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01WoHdgnDsQcrquik8wG7d6d
DoesNotAffectDryPath: at mix=0, comparing two full processor runs against each other (predelay off vs. max) isolates predelay's own contribution from the rest of the signal path - a direct comparison against the raw synthetic input is invalid, since the global input gate (applied before the chain) isn't a pure passthrough for noise. Max diff is floating-point rounding noise (~-101dB), not a leak. DelaysOnlyTheWetPath: at mix=1, confirms the wet onset shifts by exactly the requested predelay. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01WoHdgnDsQcrquik8wG7d6d
New "Pre" knob (0-1000ms) between Mix and the output rail, IR blocks only - same KnobControl pattern, drag/fine/double-click/Alt-reset behavior, and optimistic-state sync as In/Mix/Out. predelayMsScale formats the live readout as "340 ms" during drag, matching the existing gainDbScale-style scale convention. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01WoHdgnDsQcrquik8wG7d6d
Adds a 2-segment Attack/Decay envelope over an IR's detected content length (Space Designer-style), plus the content-length detection/trim it builds on: - Content length: RMS-threshold based detection of an IR's real extent (irContentLengthMs), independent of the file's raw duration. - Envelope: Init Level (level at sample 0) ramps via the Attack segment up to unity/0dB at a peak position (Attack Length, a fraction of the total trimmed length), then the Decay segment continues from that peak to the trimmed end (Decay Level). Both segments have their own Curve shape parameter. All six values travel together through setBlockIrDecay (not the per-parameter setBlockParam path) since the rebuild happens off-thread and a drag on one value can't be allowed to clobber another's in-flight value. - Rebuilds the convolver via prepareIrShapeRebuild on a background thread, same pool as model loading. Covered by ir_content_length_tests.cpp (detection sanity against real IR assets) and ir_decay_tests.cpp (truncation, dry-path isolation, tail attenuation, curve shaping, and state-restore round trip). Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0153YRkpUTsAdrZdgtbbsnLW
Front-end for the DSP envelope feature: a static waveform display for the block card, then a fully interactive drag-on-graph editor on top of it. - decayEnvelope.ts / WaveformDisplay.tsx: client-side mirror of the native envelope math for drawing the curve, plus the min/max waveform polygon and cut-line overlay. WaveformDisplay itself stays display-only (used for the small info-panel thumbnail). - useIrWaveform.ts: fetches per-block waveform peaks, refetching on both activeModelId and the modelLoading true->false edge (the latter closes a real race - see the hook's own comment). - IrEnvelopeGraph.tsx: the interactive editor replacing the waveform in the block's main view. Three draggable points (Init/Peak/End) plus draggable curve segments, each with hit-testing mirroring BlockEqView's own dot pattern (oversized invisible hit target under a small visible dot/line), hover rings, a floating value readout during drag, Shift-drag fine mode, and Alt-click/double-click reset. Includes an under-curve fill (area between the curve and the waveform below it), a faint background time grid, and a left-edge predelay indicator - all sharing the same fixed-window coordinate system the old standalone Length knob used, so dragging never rescales the backdrop. - EditableChip.tsx: extracted out of BlockEqView.tsx (zero behavior change there) so the envelope row's six precision-entry chips (Init/A Len/A Crv/D Len/D Lvl/D Crv) share the same click-to-type pattern as the EQ's Freq/Gain/Q chips - graph for by-ear shaping, chips for landing an exact value. - chain.ts / knobScale.ts / useChainActions.ts / useChainState.ts / Plugin.tsx: BlockParams fields, display scales, and the setBlockIrDecay action wiring the graph/chips commit through. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0153YRkpUTsAdrZdgtbbsnLW
…sitivity A drag of about 1/5 of the graph's height (curveNormalized moving from 0.5 to ~0.7) was already enough to push the exponent to k~2, at which point ~91% of the segment's entire dB change lands in its last 30% - close to a step function, and audibly an abrupt "choke" nothing on the coarsely-sampled graph visually suggested. Two compounding causes: the straight kCurveMax^(2*(c-0.5)) mapping has a steep slope right at center (dk/dc = 2*ln(kCurveMax) =~ 3.6), and fraction^k is applied against dB's own huge dynamic range, so even a modest k deviation from 1 concentrates almost the whole transition into a sliver of the segment. Fix: cube (curveNormalized - 0.5) before exponentiating, in both ProcessorModelLoader.cpp's shapeCurveNormalized and decayEnvelope.ts's own copy (kept in sync by hand, as before). This spreads sensitivity unevenly across the knob's travel - landing a given k near center now takes roughly double the drag distance it used to - while leaving both extremes (curveNormalized=0 and =1) mapped to exactly the same k as before, so the existing curve-shape tests (comparing Curve=0 against Curve=1) are unaffected. kCurveMax itself stays at 6.0: lowering it instead would have permanently capped how extreme the curve can go at full knob throw, not just tamed the center. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0153YRkpUTsAdrZdgtbbsnLW
…r top - Curve segments (attack/decay hit-paths) had Alt-click reset wired up but no onDoubleClick, unlike the three points (Init/Peak/End), which had both. Added, mirroring the points exactly - same resetTarget call either way. - The floating drag readout always offset above its anchor (translateY -140%). Fine everywhere except near the top of the graph (dragging Init toward unity, or Peak, which always sits there) where that pushed the chip above the card's own overflow:hidden boundary. Flips to offset below instead when the anchor is in the top 20% of the graph - standard tooltip edge-avoidance, no change to the SVG/ viewBox geometry needed since this is a plain HTML overlay. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0153YRkpUTsAdrZdgtbbsnLW
…aterial The center-sensitivity remap (previous commit) fixed how much drag it takes to move away from linear, but left the ceiling itself untouched - at full deliberate knob-throw, kCurveMax=6.0 still meant 99.93% of a segment's entire dB change landed in its last 30% (end-slope 6x steeper than linear). By ear, that's a real problem specifically for real IR material: reverb/room captures already carry their own natural decay, and the envelope's imposed dB drop *adds* to that (dB being logarithmic, multiplying amplitudes is adding dB) rather than replacing it - stacking an already-steep artificial cutoff on top of whatever the IR's own tail was doing read as a hard gate, not a usable creative curve, even at an intentional full-throw setting. kCurveMax=4.0 keeps a real, audibly-different extreme available at full throw (4x steeper than linear) while landing roughly an order of magnitude less compressed at the 30%-into-the-segment mark (0.81% vs. 0.07%). This is an independent lever from the center-sensitivity remap - one controls how much drag reaches a given curve, the other how extreme the curve can get at all - so it doesn't undo or interact with that fix, just tames the other end. Existing curve-shape tests (which compare Curve=0 against Curve=1, not against a specific k) are unaffected. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0153YRkpUTsAdrZdgtbbsnLW
Value only now ("-6dB", "240ms") instead of "Init -6dB" - less
cluttered, and the shorter text also gives the near-top edge-flip
fix (previous session) more effective margin, without needing any
further positioning-logic changes.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0153YRkpUTsAdrZdgtbbsnLW
1 task
Member
|
@Shayshez this is pretty wild will def give it a proper look, thanks! |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Adds shaping controls for IR blocks: a Predelay knob, a 2-segment Attack/Decay
envelope over the IR's detected content length (Space Designer-style), and a
fully interactive graph to shape it by ear, backed by precision-entry chips
for landing exact values.
convolver (real-time DSP stage, own knob in the Input rail).
(RMS-threshold based, independent of the file's raw duration), plus a
2-segment Attack/Decay envelope over it - Init Level ramps up to unity at a
peak position (Attack Length, a fraction of the trimmed total), then Decay
continues from that peak to the trimmed end (Decay Level). Both segments
have their own Curve shape parameter.
the IR's actual audio, auto-fit to the detected content length rather
than the file's raw duration - a 20ms cab capture with a long silent tail
displays at a legible scale instead of being squeezed into a few pixels.
The envelope graph and its interactive points/curves render directly on
top of this.
draggable curve segments directly on the waveform, mirroring the EQ
editor's own dot-dragging conventions (hover rings, Shift-drag fine mode,
Alt-click/double-click reset, a floating value readout during drag). An
under-curve fill, a faint background time grid, and a left-edge predelay
indicator round out the visualization - all within the same fixed-window
coordinate system the original Length knob used.
chips (shares its pattern with the EQ's own Freq/Gain/Q chips, extracted
into a common
EditableChip).A DSP tuning note worth flagging in review
The Curve parameter's normalized-to-exponent mapping got two passes after
by-ear testing surfaced a real oversensitivity: a drag of about 1/5 of the
graph's height was enough to push the effective exponent to a point where
~91% of a segment's entire dB change lands in its last 30% - audibly an
abrupt "choke" nothing on the graph visually suggested. Fixed by (1) cubing
the normalized input before exponentiating, which roughly doubles the drag
distance needed to reach a given curve near center while leaving both
extremes unchanged, and (2) lowering the ceiling itself (
kCurveMax6.0 -> 4.0), since real reverb/room IRs already carry their own natural
decay and the envelope's dB drop adds to that rather than replacing it -
even the old ceiling's intentional full-throw extreme read as a hard gate
once stacked on real material. Both changes are native/TS pairs kept in
sync by hand (
ProcessorModelLoader.cpp/decayEnvelope.ts), and existingcurve-shape tests are unaffected since they compare against the (unchanged)
extremes.
Test plan
npm run lint/npm run build(UI) - clean./script/test-dsp.sh- 150/150 tests pass, including newIrContentLengthTestandIrDecayTestcoverage (truncation, dry-pathisolation, tail attenuation, curve shaping, state-restore round trip)
🤖 Generated with Claude Code