[Free Dsp] Lookahead Limiter (true peak)

[griffinboy]

Lookahead Limiter designed for low distortion.
It's not as fancy as Pro-L but it does the trick.
Oversample it for True Peak limiting.

Stereo Only
CPU: (0.2%)

*Important warning: DAW latency correction!
The latency changes depending on attack time in this script, but in hise it's common to report a single latency value to the host. The solution to this is to either report the max latency to the host (measure from the max attack time), or to leave the attack time fixed, so that you can report a single latency value that won't change. You should be fine leaving it at 1 - 10 ms. A bit of attack (lookahead) is needed for smooth limiting.

Download the script here:

https://1drv.ms/u/c/6c19b197e5968b36/EQFgIcgl83VJviZoznFZWnEBleCXFfAD-KejdhTV9Q6IwQ?e=EyH8fM

Setup tutorial:

1. Create a 3rd party C++ node in HISE named "Griffin_Limiter"
2. Compile the initial DLL in HISE using "Compile dsp networks as .dll"
3. Replace the generated "Griffin_Limiter.h" (located in ProjectName/DspNetworks/ThirdParty) with my limiter header file
4. Re-compile the final DLL in HISE using "Compile dsp networks as .dll"

Ignore the audio artefacts in the video. My encoder is being a fool. The sound quality of the dsp is fine.

You can read the script here:
Griffin_Limiter.h:

#pragma once
#include <JuceHeader.h>
#include <cstdint>
#include <cmath>
#include <vector>
#include <algorithm>
#include <limits>
#include <cstring>

/**

==============| by GriffinBoy (2025) |==============

This node implements a Lookahead Stereo Limiter

Features:
- ExpSmootherCascade: Four-stage exponential smoothing with separate attack and release coefficients.
- PeakHoldCascade: Eight-stage peak detection cascade for accurate envelope following.
- Integrated lookahead
- Parameter smoothing 

Integration Steps for HISE:
1. Create a 3rd party C++ node in HISE named " Griffin_Limiter "
2. Compile the initial DLL in HISE using "Compile dsp networks as .dll"
3. Replace the generated "Griffin_Limiter.h" (located in ProjectName/DspNetworks/ThirdParty) with this header file you are reading now
4. Re-compile the final DLL in HISE using "Compile dsp networks as .dll"

Author: GriffinBoy (2025)
License: UwU

Based on the following paper:
Sanfilippo, D. (2022).Envelope following via cascaded exponential smoothers for low - distortion peak limiting and maximisation.In Proceedings of the International Faust Conference, Saint - Étienne, France.

*/


namespace project
{

#ifndef M_PI
#define M_PI 3.14159265358979323846
#endif

    using namespace juce;
    using namespace hise;
    using namespace scriptnode;

    namespace FunctionsClasses {

        //------------------------------------------------------------------------------
        // DelaySmooth: A delay line with smooth crossfading between two integer delay
        // lines to produce a click-free, Doppler-free delay variation.
        // Fixed types: head = uint16_t and sample type = float.
        class DelaySmooth {
        private:
            const size_t bufferLen = 2ul << (8 * sizeof(uint16_t) - 1); 
            size_t delay = 0;
            size_t interpolationTime = 1024;
            size_t lowerDelay = 0;
            size_t upperDelay = 0;
            float interpolation = 0.0f;
            float interpolationStep = 1.0f / float(interpolationTime);
            float increment = interpolationStep;
            uint16_t lowerReadPtr = 0;
            uint16_t upperReadPtr = 0;
            uint16_t writePtr = 0;
            std::vector<float> bufferLeft;
            std::vector<float> bufferRight;
        public:
            void SetDelay(size_t _delay) { delay = _delay; }
            void SetInterpolationTime(size_t _interpTime) {
                interpolationTime = std::max<size_t>(1, _interpTime);
                interpolationStep = 1.0f / float(interpolationTime);
            }
            void Reset() {
                std::fill(bufferLeft.begin(), bufferLeft.end(), 0.0f);
                std::fill(bufferRight.begin(), bufferRight.end(), 0.0f);
            }
            void Process(float** xVec, float** yVec, size_t vecLen) {
                float* xLeft = xVec[0];
                float* xRight = xVec[1];
                float* yLeft = yVec[0];
                float* yRight = yVec[1];
                for (size_t n = 0; n < vecLen; n++) {
                    bufferLeft[writePtr] = xLeft[n];
                    bufferRight[writePtr] = xRight[n];
                    bool lowerReach = (interpolation == 0.0f);
                    bool upperReach = (interpolation == 1.0f);
                    bool lowerDelayChanged = (delay != lowerDelay);
                    bool upperDelayChanged = (delay != upperDelay);
                    bool startDownwardInterp = (upperReach && upperDelayChanged);
                    bool startUpwardInterp = (lowerReach && lowerDelayChanged);
                    float incrementPathsUp[2] = { increment, interpolationStep };
                    float incrementPathsDown[2] = { incrementPathsUp[startUpwardInterp], -interpolationStep };
                    increment = incrementPathsDown[startDownwardInterp];
                    size_t lowerDelayPaths[2] = { lowerDelay, delay };
                    size_t upperDelayPaths[2] = { upperDelay, delay };
                    lowerDelay = lowerDelayPaths[upperReach];
                    upperDelay = upperDelayPaths[lowerReach];
                    // Use explicit bitwise AND for modulo (bufferLen = 65536)
                    lowerReadPtr = static_cast<uint16_t>(writePtr - lowerDelay) & 0xFFFF;
                    upperReadPtr = static_cast<uint16_t>(writePtr - upperDelay) & 0xFFFF;
                    writePtr++;
                    interpolation = std::max(0.0f, std::min(1.0f, interpolation + increment));
                    yLeft[n] = interpolation * (bufferLeft[upperReadPtr] - bufferLeft[lowerReadPtr]) + bufferLeft[lowerReadPtr];
                    yRight[n] = interpolation * (bufferRight[upperReadPtr] - bufferRight[lowerReadPtr]) + bufferRight[lowerReadPtr];
                }
            }
            DelaySmooth() {
                bufferLeft.resize(bufferLen);
                bufferRight.resize(bufferLen);
            }
            DelaySmooth(size_t _delay, size_t _interpTime) {
                bufferLeft.resize(bufferLen);
                bufferRight.resize(bufferLen);
                delay = _delay;
                interpolationTime = std::max<size_t>(1, _interpTime);
                interpolationStep = 1.0f / float(interpolationTime);
            }
        };

        //------------------------------------------------------------------------------
        // ExpSmootherCascade: Cascaded one-pole exponential smoothers (4 stages)
        // with separate attack and release coefficients.
        class ExpSmootherCascade {
        private:
            static constexpr size_t stages = 4;
            const float coeffCorrection = 1.0f / std::sqrt(std::pow(2.0f, 1.0f / float(stages)) - 1.0f);
            const float epsilon = std::numeric_limits<float>::epsilon();
            float SR = 48000.0f;
            float T = 1.0f / SR;
            const float twoPiC = 2.0f * static_cast<float>(M_PI) * coeffCorrection;
            float twoPiCT = twoPiC * T;
            float attTime = 0.001f;
            float relTime = 0.01f;
            float attCoeff = std::exp(-twoPiCT / attTime);
            float relCoeff = std::exp(-twoPiCT / relTime);
            float coeff[2] = { relCoeff, attCoeff };
            float output[stages] = { 0.0f, 0.0f, 0.0f, 0.0f };
        public:
            void SetSR(float _SR) {
                SR = std::max(_SR, 1.0f);
                T = 1.0f / SR;
                twoPiCT = twoPiC * T;
            }
            void SetAttTime(float _attTime) {
                attTime = std::max(epsilon, _attTime);
                attCoeff = std::exp(-twoPiCT / attTime);
                coeff[1] = attCoeff;
            }
            void SetRelTime(float _relTime) {
                relTime = std::max(epsilon, _relTime);
                relCoeff = std::exp(-twoPiCT / relTime);
                coeff[0] = relCoeff;
            }
            void Reset() { std::memset(output, 0, sizeof(output)); }
            void Process(float* xVec, float* yVec, size_t vecLen) {
                for (size_t n = 0; n < vecLen; n++) {
                    float input = xVec[n];
                    // Unrolled stage 0
                    bool isAttackPhase = (input > output[0]);
                    float coeffVal = isAttackPhase ? attCoeff : relCoeff;
                    output[0] = input + coeffVal * (output[0] - input);
                    input = output[0];
                    // Unrolled stage 1
                    isAttackPhase = (input > output[1]);
                    coeffVal = isAttackPhase ? attCoeff : relCoeff;
                    output[1] = input + coeffVal * (output[1] - input);
                    input = output[1];
                    // Unrolled stage 2
                    isAttackPhase = (input > output[2]);
                    coeffVal = isAttackPhase ? attCoeff : relCoeff;
                    output[2] = input + coeffVal * (output[2] - input);
                    input = output[2];
                    // Unrolled stage 3
                    isAttackPhase = (input > output[3]);
                    coeffVal = isAttackPhase ? attCoeff : relCoeff;
                    output[3] = input + coeffVal * (output[3] - input);
                    yVec[n] = output[3];
                }
            }
            ExpSmootherCascade() {}
            ExpSmootherCascade(float _SR, float _attTime, float _relTime) {
                SR = std::max(_SR, 1.0f);
                T = 1.0f / SR;
                twoPiCT = twoPiC * T;
                attTime = std::max(epsilon, _attTime);
                relTime = std::max(epsilon, _relTime);
                attCoeff = std::exp(-twoPiCT / attTime);
                relCoeff = std::exp(-twoPiCT / relTime);
                coeff[0] = relCoeff;
                coeff[1] = attCoeff;
            }
        };

        //------------------------------------------------------------------------------
        // PeakHoldCascade: Cascaded peak-holders (8 stages fixed) to approximate a max filter.
        // The cascade splits the hold time to detect secondary peaks.
        class PeakHoldCascade {
        private:
            static constexpr size_t stages = 8;
            float SR = 48000.0f;
            float holdTime = 0.001f;
            const float oneOverStages = 1.0f / float(stages);
            size_t holdTimeSamples = std::rint(holdTime * oneOverStages * SR);
            size_t timer[stages];
            float output[stages];
        public:
            void SetSR(float _SR) {
                SR = std::max(_SR, 1.0f);
                holdTimeSamples = std::rint(holdTime * oneOverStages * SR);
            }
            void SetHoldTime(float _holdTime) {
                holdTime = std::max(0.0f, _holdTime);
                holdTimeSamples = std::rint(holdTime * oneOverStages * SR);
            }
            void Reset() {
                std::memset(timer, 0, sizeof(timer));
                std::memset(output, 0, sizeof(output));
            }
            void Process(float* xVec, float* yVec, size_t vecLen) {
                for (size_t n = 0; n < vecLen; n++) {
                    float input = std::fabs(xVec[n]);
                    bool release;
                    // Unrolled stage 0
                    release = (input >= output[0]) || (timer[0] >= holdTimeSamples);
                    timer[0] = release ? 0 : (timer[0] + 1);
                    output[0] = release ? input : output[0];
                    input = output[0];
                    // Unrolled stage 1
                    release = (input >= output[1]) || (timer[1] >= holdTimeSamples);
                    timer[1] = release ? 0 : (timer[1] + 1);
                    output[1] = release ? input : output[1];
                    input = output[1];
                    // Unrolled stage 2
                    release = (input >= output[2]) || (timer[2] >= holdTimeSamples);
                    timer[2] = release ? 0 : (timer[2] + 1);
                    output[2] = release ? input : output[2];
                    input = output[2];
                    // Unrolled stage 3
                    release = (input >= output[3]) || (timer[3] >= holdTimeSamples);
                    timer[3] = release ? 0 : (timer[3] + 1);
                    output[3] = release ? input : output[3];
                    input = output[3];
                    // Unrolled stage 4
                    release = (input >= output[4]) || (timer[4] >= holdTimeSamples);
                    timer[4] = release ? 0 : (timer[4] + 1);
                    output[4] = release ? input : output[4];
                    input = output[4];
                    // Unrolled stage 5
                    release = (input >= output[5]) || (timer[5] >= holdTimeSamples);
                    timer[5] = release ? 0 : (timer[5] + 1);
                    output[5] = release ? input : output[5];
                    input = output[5];
                    // Unrolled stage 6
                    release = (input >= output[6]) || (timer[6] >= holdTimeSamples);
                    timer[6] = release ? 0 : (timer[6] + 1);
                    output[6] = release ? input : output[6];
                    input = output[6];
                    // Unrolled stage 7
                    release = (input >= output[7]) || (timer[7] >= holdTimeSamples);
                    timer[7] = release ? 0 : (timer[7] + 1);
                    output[7] = release ? input : output[7];
                    yVec[n] = output[7];
                }
            }
            PeakHoldCascade() { Reset(); }
            PeakHoldCascade(float _SR, float _holdTime) {
                SR = std::max(_SR, 1.0f);
                holdTime = _holdTime;
                holdTimeSamples = std::rint(holdTime * oneOverStages * SR);
                Reset();
            }
        };

        //------------------------------------------------------------------------------
        // Limiter: Lookahead peak-limiter combining DelaySmooth, PeakHoldCascade, and
        // ExpSmootherCascade. Computes a stereo envelope and applies a gain to keep
        // signal peaks below a given threshold.
        template<typename real>
        class Limiter {
        private:
            float SR = 48000.0f;
            float T = 1.0f / SR;
            const float twoPi = 2.0f * static_cast<float>(M_PI);
            const float epsilon = std::numeric_limits<float>::epsilon();
            const float smoothParamCutoff = 20.0f;
            float attack = 0.01f;
            float hold = 0.0f;
            float release = 0.05f;
            float dBThreshold = -6.0f;
            float linThreshold = std::pow(10.0f, dBThreshold * 0.05f);
            float dBPreGain = 0.0f;
            float linPreGain = 1.0f;
            float smoothPreGain = 0.0f;
            float smoothThreshold = 0.0f;
            float smoothParamCoeff = std::exp(-twoPi * smoothParamCutoff * T);
            size_t lookaheadDelay = 0;
            DelaySmooth delay;
            static constexpr size_t numberOfPeakHoldSections = 8;
            static constexpr size_t numberOfSmoothSections = 4;
            const float oneOverPeakSections = 1.0f / float(numberOfPeakHoldSections);
            PeakHoldCascade peakHolder;
            ExpSmootherCascade expSmoother;
        public:
            void SetSR(float _SR) {
                SR = std::max(_SR, 1.0f);
                T = 1.0f / SR;
                smoothParamCoeff = std::exp(-twoPi * smoothParamCutoff * T);
                peakHolder.SetSR(SR);
                expSmoother.SetSR(SR);
            }
            void SetAttTime(float _attack) {
                attack = std::max(epsilon, _attack);
                lookaheadDelay = std::rint(attack * oneOverPeakSections * SR) * numberOfPeakHoldSections;
                delay.SetDelay(lookaheadDelay);
                delay.SetInterpolationTime(lookaheadDelay);
                expSmoother.SetAttTime(attack);
                peakHolder.SetHoldTime(attack + hold);
            }
            void SetHoldTime(float _hold) {
                hold = std::max(0.0f, _hold);
                peakHolder.SetHoldTime(attack + hold);
            }
            void SetRelTime(float _release) {
                release = std::max(epsilon, _release);
                expSmoother.SetRelTime(release);
            }
            void SetThreshold(float _threshold) {
                dBThreshold = std::max(-120.0f, _threshold);
                linThreshold = std::pow(10.0f, dBThreshold * 0.05f);
            }
            void SetPreGain(float _preGain) {
                dBPreGain = _preGain;
                linPreGain = std::pow(10.0f, dBPreGain * 0.05f);
            }
            void Reset() {
                delay.Reset();
                peakHolder.Reset();
                expSmoother.Reset();
            }
            // Process takes separate input and output stereo buffers.
            // The design uses an in-place delay so the input buffer is overwritten.
            void Process(float** xVec, float** yVec, size_t vecLen) {
                // Get channel pointers once outside loops.
                float* xLeft = xVec[0];
                float* xRight = xVec[1];
                float* yLeft = yVec[0];
                float* yRight = yVec[1];
                // Merge pre-gain smoothing and envelope computation.
                for (size_t n = 0; n < vecLen; n++) {
                    smoothPreGain = linPreGain + smoothParamCoeff * (smoothPreGain - linPreGain);
                    xLeft[n] *= smoothPreGain;
                    xRight[n] *= smoothPreGain;
                    yLeft[n] = std::max(std::fabs(xLeft[n]), std::fabs(xRight[n]));
                }
                // Process envelope with peak-hold cascade.
                peakHolder.Process(yLeft, yLeft, vecLen);
                // Smooth and clip envelope to threshold.
                for (size_t n = 0; n < vecLen; n++) {
                    smoothThreshold = linThreshold + smoothParamCoeff * (smoothThreshold - linThreshold);
                    yLeft[n] = std::max(yLeft[n], smoothThreshold);
                    yRight[n] = smoothThreshold;
                }
                // Smooth envelope with exponential cascade.
                expSmoother.Process(yLeft, yLeft, vecLen);
                // Apply lookahead delay (in-place).
                delay.Process(xVec, xVec, vecLen);
                // Compute attenuation gain and apply to delayed signal in one loop.
                for (size_t n = 0; n < vecLen; n++) {
                    float gain = yRight[n] / yLeft[n];
                    yLeft[n] = gain * xLeft[n];
                    yRight[n] = gain * xRight[n];
                }
            }
            Limiter() {}
            Limiter(float _SR, float _dBPreGain, float _attack, float _hold, float _release, float _dBThreshold) {
                SR = std::max(_SR, 1.0f);
                dBPreGain = _dBPreGain;
                attack = std::max(epsilon, _attack);
                hold = std::max(0.0f, _hold);
                release = std::max(epsilon, _release);
                dBThreshold = std::max(-120.0f, _dBThreshold);
            }
        };

    } // end namespace FunctionsClasses

    //------------------------------------------------------------------------------
    // SNEX Node - Stereo Limiter Node Implementation
    //------------------------------------------------------------------------------
    template <int NV>
    struct Griffin_Limiter : public data::base
    {
        SNEX_NODE(Griffin_Limiter);

        struct MetadataClass
        {
            SN_NODE_ID("Griffin_Limiter");
        };

        // Node Properties 
        static constexpr bool isModNode() { return false; }
        static constexpr bool isPolyphonic() { return NV > 1; }
        static constexpr bool hasTail() { return false; }
        static constexpr bool isSuspendedOnSilence() { return false; }
        static constexpr int getFixChannelAmount() { return 2; }

        static constexpr int NumTables = 0;
        static constexpr int NumSliderPacks = 0;
        static constexpr int NumAudioFiles = 0;
        static constexpr int NumFilters = 0;
        static constexpr int NumDisplayBuffers = 0;

        // Create an instance of our DSP Limiter 
        FunctionsClasses::Limiter<float> limiter;

        // Scratch buffers to avoid per-block allocation.
    private:
        std::vector<float> scratchInLeft, scratchInRight;
        std::vector<float> scratchOutLeft, scratchOutRight;
    public:
        //--------------------------------------------------------------------------
        // Main Processing Functions
        //--------------------------------------------------------------------------
        void prepare(PrepareSpecs specs)
        {
            float sampleRate = specs.sampleRate;
            limiter.SetSR(sampleRate);
            limiter.Reset();
            // Preallocate scratch buffers (use maximum BlockSize if available, otherwise default to 512).
            int blockSize = (specs.blockSize > 0) ? specs.blockSize : 512;
            scratchInLeft.resize(blockSize);
            scratchInRight.resize(blockSize);
            scratchOutLeft.resize(blockSize);
            scratchOutRight.resize(blockSize);
        }

        void reset() {}

        template <typename ProcessDataType>
        inline void process(ProcessDataType& data)
        {
            auto& fixData = data.template as<ProcessData<getFixChannelAmount()>>();
            auto audioBlock = fixData.toAudioBlock();
            float* leftChannelData = audioBlock.getChannelPointer(0);
            float* rightChannelData = audioBlock.getChannelPointer(1);
            int numSamples = data.getNumSamples();

            juce::FloatVectorOperations::copy(scratchInLeft.data(), leftChannelData, numSamples);
            juce::FloatVectorOperations::copy(scratchInRight.data(), rightChannelData, numSamples);

            float* inBuffers[2] = { scratchInLeft.data(), scratchInRight.data() };
            float* outBuffers[2] = { scratchOutLeft.data(), scratchOutRight.data() };

            limiter.Process(inBuffers, outBuffers, numSamples);

            juce::FloatVectorOperations::copy(leftChannelData, scratchOutLeft.data(), numSamples);
            juce::FloatVectorOperations::copy(rightChannelData, scratchOutRight.data(), numSamples);
        }

        //--------------------------------------------------------------------------
        // Parameter Handling
        //--------------------------------------------------------------------------
        template <int P>
        void setParameter(double v)
        {
            if (P == 0) {
                limiter.SetPreGain(static_cast<float>(v));
            }
            else if (P == 1) {
                // Convert from ms to seconds.
                limiter.SetAttTime(static_cast<float>(v * 0.001));
            }
            else if (P == 2) {
                limiter.SetHoldTime(static_cast<float>(v * 0.001));
            }
            else if (P == 3) {
                limiter.SetRelTime(static_cast<float>(v * 0.001));
            }
            else if (P == 4) {
                limiter.SetThreshold(static_cast<float>(v));
            }
        }

        void createParameters(ParameterDataList& data)
        {
            {
                parameter::data p("PreGain (dB)", { -24.0, 24.0, 0.1 });
                registerCallback<0>(p);
                p.setDefaultValue(0.0);
                data.add(std::move(p));
            }
            {
                parameter::data p("Attack (ms)", { 1.0, 500.0, 1.0 });
                registerCallback<1>(p);
                p.setDefaultValue(10.0);
                p.setSkewForCentre(80.0f);
                data.add(std::move(p));
            }
            {
                parameter::data p("Hold (ms)", { 0.0, 100.0, 1.0 });
                registerCallback<2>(p);
                p.setDefaultValue(5.0);
                p.setSkewForCentre(30.0f);
                data.add(std::move(p));
            }
            {
                parameter::data p("Release (ms)", { 1.0, 2500.0, 1.0 });
                registerCallback<3>(p);
                p.setDefaultValue(80.0);
                p.setSkewForCentre(800.0f);
                data.add(std::move(p));
            }
            {
                parameter::data p("Ceiling (dB)", { -60.0, 0.0, 0.1 });
                registerCallback<4>(p);
                p.setDefaultValue(-6.0);
                data.add(std::move(p));
            }
        }

        void setExternalData(const ExternalData& ed, int index)
        {
            // Not needed.
        }

        void handleHiseEvent(HiseEvent& e)
        {
            // Not needed.
        }

        template <typename FrameDataType>
        void processFrame(FrameDataType& data)
        {
            // Not needed.
        }
    };

} // end namespace project

[ustk]

@griffinboy Cool stuffs!
But I never fully understood the meaning of true peak in the context of a dynamic effect.
The minimal attack time here is 1ms. So this means during this time the signal could reach any value, right? So why true peak if this exceeding part can't be cut? An I mixing things in my head?

[griffinboy]

@ustk

Haha it took me a long time to understand.

True peak is needed because the waveform later down the line (when converted to analog) can have peaks inbetween the digital samples, that are higher than those found in the samples. True peak detection uses oversampling or interpolation to predict the in-between peaks. In a limiter, this means the compressor can detect and reduce peaks that might cause distortion or clip after converted to analog, even if the digital samples look safe. True Peak causes the limiter to create a different gain reduction response, to catch these invisible peaks, which does affect the sound. Some people don't like the sound that this imposes.

As for lookahead, we use latency to give the limiter a head start to begin reducing the gain before a transient arrives.
We need to do this because, in a limiter, we smooth the gain reduction, to avoid distortion that we'd get if we yanked the signal down hard. But this smoothing adds extra time for the limiter to reach full reduction after detecting a transient , so we have to start reducing gain before the transient arrives.

Don't be fooled by attack time here... In this design, when you add more attack time, we also add more lookahead, so we still end up grabbing the peaks correctly (peaks don't escape the attack), but rather the smoothing curve is changed during the attack phase, which will affect the groove.

[griffinboy]

This post is deleted!

[ustk]

@griffinboy Oh I get it now! I understand how true peak is working and the signal in-between sample can reach different values once upsampled or converted back to analog.

Those peaks are then read to set the appropriate reduction, it makes sense. But then only in a lookahead algorithm, right? Otherwise they'll just pass through...

So it's simple in the end, it's just my mind trying to make everything more complex, the story of my life

[griffinboy]

@ustk

well yes, you're only meant to upsample the detection signal not the real signal (I think? But I may be wrong actually). But I haven't tried using hise oversampling in c++ yet so I was lazy and just said to wrap the whole node in oversampling, which is bad for performance, but I'll look at doing it properly at somepoint lol. It was only whipped up in a day to help another user

[orange]

@griffinboy Is there a fix for the latency that the lookahead limiter introduces?

Delay compensation must be a static value.

[griffinboy]

@orange

Good point. I'm not sure how hise measures this. Because there is a function for it, but idk if it will count custom nodes, or whether I have to do something in my nodes for this?

@Christoph-Hart ?

[orange]

@griffinboy Normally, a declaration is made with using Engine.setLatencySamples() so that the DAW knows how much is the the latency of the plugin on init.

You measure the latency produced by the plugin empirically and use that value. And this value is a couple of samples normally. But the lookahead limiter also changes according to the latency attack value, and the latency is high, right? The stock limiter in Hise is like this too.

[griffinboy]

@orange

Ah I get your point! Indeed. Perhaps I should've released this with a fixed attack - I'm not sure anyone really needs to change it all that much.

It's fixed in many limiter plugins, and software like pro-l uses the attack parameter for a different purpose anyway.

I'm not sure there's an easy way to make sure the latency compensation is right with what I released. Thanks for bringing it up.

I guess the safest way at the moment would be to assume max latency, or leave the attack fixed. I've included a warning in my original post now.

I wonder how other plugins solve this. This is essentially an issue of variable lookahead time.

[ustk]

@griffinboy said in [Free Dsp] Lookahead Limiter (true peak):

well yes, you're only meant to upsample the detection signal not the real signal

If so, then I think the detection signal will be delayed in regards to the real signal because of the oversampling. Resulting in compression not applied at the exact moment (probably a few samples later) but I might be wrong.

What I know is that in the plugin I am actually releasing, I had to compensate some branched signal for the ones that are using oversampling, otherwise they do not arrive in phase at the end of the chain. The issue might be less problematic in a compressor since you don't blend the two signals, but...

@orange is right, this is how one would set the latency, emprirical measures and report with that function.
A static latency is better, but I still change mine in regards with the OS factor which does not happen at a "tweak the knobs" level. So for an attack knob, it should be fixed IMHO

[griffinboy]

@ustk

It wouldn't be hard to map the oversampled reduction to the lower sample rate original signal. I'll probably revisit this at somepoint and do it.

[ustk]

@griffinboy said in [Free Dsp] Lookahead Limiter (true peak):

I wonder how other plugins solve this. This is essentially an issue of variable lookahead time.

probably just something like

MAX_LOOKAHEAD = 48000;

Engine.setLatencySamples(MAX_LOOKAHEAD);

delayCompensation = MAX_LOOKAHEAD - attackTime;

[griffinboy]

@ustk

hmm maybe I should ask for Hise support for variable host latency.
But I think this subject is fiddly enough, Christoph has already done plenty of work trying to get latency to be consistent in the hosts.

[ustk]

@griffinboy For what I understand, it's not Hise to support latency, be it variable or not. Hise just reports what you want, so if you want it to be variable, technically, you should be able to do it.
But the DAW might not like it... It doesn't seem to be something "normal" to report a continuously changing value, but I can't refer to anything else than my own perfectible thoughts...