Almost fifteen years ago I first heard about an invention that has never let go of me since. Its creator, Jens Groh, whom I have known for many years, was working at the time as a DSP developer at the Institut für Rundfunktechnik in Munich, the joint research institute of the German public broadcasters. What he had built there sounded like one of those rare cases where somebody does not work around a fundamental problem of sound recording but actually solves it: a way of summing microphone signals that makes the comb filters disappear.
Anyone who records the same source with more than one microphone knows the problem. Each microphone hears the source with a slightly different delay, and when the signals are mixed together, some frequency components add up while others cancel. The result is the comb filter effect: a frequency response shaped like a comb, a hole at 100 hertz here, a bump at 300 there, depending on distance, depending on how the musician moves, changing all the time. Recording engineers describe the result with words like “phasey,” “muddy,” “smeared.” The usual remedies are discipline in microphone placement, delay alignment, polarity guesswork and corrective EQ, and none of them solves the problem; they merely move it around.
Groh’s idea goes deeper. Instead of simply adding the signals (by voltage, which is what every mixing console does), his method continuously adjusts the sum for each frequency component so that the energy at the output comes out right. Where the channels are about to cancel each other, the sum is raised; where they are about to pile up, it is lowered. No delay alignment, no phase trickery: the correction follows the signal. The IRT described all of this between 2009 and 2017 in a whole family of patents, presented it in 2010 under the name Acoustic Summing at the Tonmeistertagung, the convention of the German association of recording engineers, and had a working real-time prototype running on dedicated hardware. The announcement text of that talk also states what the method is really about, and it is not the repair argument (my translation):
“To avoid audible degradation, microphone placement has always had to be guided, to no small degree, by avoiding this effect. This, however, puts narrow limits on the possibilities of sound design.”
So the promise is a double one: less coloration, and above all the freedom to place microphones where you want them for the sound, rather than where the interference happens to permit.
And then: nothing came of it. No product, no licensee, no quiet acquisition by one of the big manufacturers. That question stayed with me all these years. Such an elegant solution to such a ubiquitous problem, and nobody reaches for it?
The First Attempt
When the IRT was wound down, around 2021, I thought: now. Perhaps these patents could be picked up cheaply, and then one could finally experiment with them. It turned out to be surprisingly difficult to find anyone who was even responsible for the patent portfolio. I made phone calls that led nowhere. Above all, the decisive piece was missing: the practical implementation. The code from back then belongs to the IRT, which has not released it to this day, and without a working implementation a patent document is of little use to someone like me. With that, the foundation was gone. The subject went to sleep.
The Second Attempt
What woke it up again was, of all things, a different project. At CORPUS we have long been working with agentic AI systems that help us with our engineering, but that had always happened at the web level: databases, pipelines, interfaces. Signal processing at the DSP level was something I simply did not trust the models with, more out of intuition than experience. That changed radically during the work on our quality control for incoming music uploads, where, together with Claude Fable, we built a whole battery of audio detectors: clicks, dropouts, hum, hidden format conversions.
Two things were special about this work. The first is the working method: from day one we built automated evaluation loops, prepared test cases and measurement scripts against which the agent checks its own results before reporting success to me. Nothing got integrated that had not survived these gates and then proven itself in my own tests, which is where most of the iteration loops came from. The second is the true art of the trade, the false positives, because music is full of events that look like defects. A flamenco track pushed the click detector to its limits with its palmas, the rhythmic handclaps (the solution: an acoustic attack is a jump in energy, a click is not). And when four “confirmed clicks” remained on an electronic track, annoying me, Claude came up with an idea I had never seen formulated anywhere in pro audio: sample the waveform of the suspicious click and check whether the same shape recurs musically across the track.
Three of the four events were identical 8-millisecond waveforms, repeated twenty-six times on a tempo grid: sampler hits. The principle behind it has been in our lab notebook ever since: media damage is a unique accident and never repeats its waveform; whatever recurs is, by construction, music. Elsewhere, Claude pulled a technique straight over from image forensics, detecting covert sample rate conversions via the periodicity of the interpolation error, a classic of image forgery detection, transferred to audio. It worked outrageously well. And it made me want more, because suddenly someone like me, a composer with technical interests but no C++ past, had entirely new ways of building things. And I want to build things.
So I had Claude research the IRT patents, all thirteen families, in full text. The question was: is the mathematics described there in enough detail to build a prototype of our own, without the original implementation that was never released? Together with further material that surfaced during the research (including, and this became important later, a diploma thesis from 2013), it added up to a solid foundation. The core formula is in the patents in full: the target is the power sum of the inputs, plus a compensation degree D and a limiting parameter L that caps the maximum correction. Multichannel works as a cascade of two-input stages.
Along the way, the patent question resolved itself in the most astonishing manner: the register search showed that the IRT, in liquidation, had simply let the decisive rights lapse. The German and European members of the relevant patent families have expired for non-payment of the renewal fees; in Europe the method is free. All those phone calls back then, and in the end the answer was: it no longer belongs to anyone who wants to defend it.
Ten Days in August
On August 15 I started building an offline prototype in Python with Claude. The first proof of concept stood that same weekend, and the first synthetic test electrified me: white noise plus a copy delayed by 3.7 milliseconds, the textbook comb filter. The normal sum shows a ripple of almost 30 dB from peak to notch. The corrected sum: 0.2 dB. The comb filter is simply gone.
Then real music: my own drum sessions, a string trio, later an orchestra recording with 21 tracks. And here the real work began, because block-based correction creates artifacts of its own. Transients smear, because the correction wants to be faster than the signal; on sustained orchestral chords a fine flutter appears, which I named warbling. Countermeasures followed in quick succession, and some of them I still find beautiful: a transient protector that uses a harmonic-percussive decomposition to take only the percussive parts out of the correction, bin by bin instead of broadband; a smoothing that only steps in when no onset is pending. Every iteration was auditioned, compared, logged. v7, v8, v9. It felt like steady progress: the artifacts got quieter, the renderings got more pleasant.
What I failed to notice: along with the artifacts, the effect got quieter too. The smoothing and protection iterations were ironing out precisely what the method was supposed to do. A later measurement put it plainly: the transient protector pushed away the fill-in at the kick drum while the attenuations remained, so the net result at the critical spot was a deeper hole than with no correction at all. But I was not there yet. I heard improvements because I expected improvements.
The Plugin
In parallel with the work on the sound, the Python prototype grew into a real VST3 plugin, built with JUCE and CMake. I carried the working method over from the quality control one to one: here too, automated evaluation loops existed from the start, null tests in which the plugin has to prove that it computes the same result as the Python prototype down to rounding noise, and a scripted Reaper acceptance run that measures the entire chain after every change. The agent does not just build; it verifies its own work. The architecture is more unusual than that of a normal effect: a decomb instance sits on every microphone track, the instances find each other through a group logic, one of them automatically becomes the leader, computes the cascade across all tracks and distributes the correction factors back; each instance applies its own share, and the perfectly ordinary sum in the DAW then yields the comb-filter-free mix. No routing changes, no summing bus; the mixing desk stays as it is.
Along the way there was detective work of the kind every audio plugin project probably knows, except that here it happened in dialogue with an AI: the discovery that Reaper’s offline bounce processes sibling tracks offset against each other by many blocks, which dictates the latency reserve; a scheduling riddle in which the engine on a 16-core laptop seemed to run twenty times slower than in the studio, until it turned out that Reaper’s anticipative FX processing was occupying all the cores and the plugin’s worker threads were simply starving (the fix: lift our own threads into the same Windows scheduling class as the host’s audio threads); hand-generated median networks in AVX2 that made the most expensive part of the computation ten times faster.
My favorite find, though, is a different one, and it comes from exactly those evaluation loops: during an interim measurement, Claude triumphantly reported 92 percent engine share in live playback. Only a later fix to its own measurement script revealed that the play cursor of the test session had been sitting at the end of the project; what had been measured was forty seconds of silence after the material, and on silence the algorithm is dirt cheap. Claude found the error itself, owned it itself, and withdrew the number. Measuring, too, is an activity in which you find what you want to find; that goes for agents just as much as for humans, and it is the same mechanism that was about to catch up with me at the listening end.
Today the plugin works beautifully: null-test clean against the offline reference down below minus 120 dB, an orchestra of 21 tracks runs in real time on a laptop with plenty of headroom, plus a live visualization that shows the interference in the sum as a map, red for cancellation, blue for buildup.
And here I have to write down what is, for me, the most astonishing sentence of this text: this plugin, in its current state, was built in four working days. The complete path from the first line of Python to the blind-tested plugin: ten calendar days, of which, according to the lab notebook, only seven were working days at all, one of them pure patent research. I am convinced that each of these days stands for at least one developer month in the old world, and I do not think I am exaggerating. Close to 12,000 lines of code were written in those days, a good 5,900 of them in the C++ plugin, and I do not know a single one of them: in all that time I never once looked at the code. This is not a marginal improvement in engineering productivity. This is another world.
The Reckoning
The longer I listened, the louder a quiet suspicion became that I had ignored at first. The v9 demos I sent to befriended recording engineers sounded good, but the differences from the normal sum seemed to shrink the more often I compared. So I built what I should have built from the start: a double-blind test. A script generated the project’s entire listening history as ABX pairs, sample-aligned, level-matched, thirty pairs from the very first noise test to the current state, auditioned with the ABX comparator in foobar2000.
The noise: eight out of eight correct; by pure guessing you manage that with a chance of one in 256. The test chain works, and so does my hearing.
The music: chance level. The cascade on the drums, whose superiority I had heard sighted as “clear” just one week earlier (“more body in snare and kick, the plain sum sounds cardboardy,” says the log): five out of eight. The most aggressive stage of the method on the most forgiving material, the string trio: five out of eight. Even the warbling, the artifact against which I had had a whole generation of countermeasures built, did not survive the blind test at this spot. The only thing that initially did hold up blind was, of all things, the damage done by the maximum setting: with the correction fully open, the kick drum turns audibly soft and flabby, eight out of eight, identified as an artifact, not as a benefit.
That was the painful moment. And in hindsight I can see exactly how I had maneuvered myself into it: the noise example was so spectacularly unambiguous that it set my expectation for everything that followed. If the effect is this dramatic in the textbook case, surely it must be there in music too. So I heard it. I have worked with my ears for decades, I know that hearing is a subjective process, I have run into “you hear what you want to hear” again and again over my career, in myself and in others. It got me anyway. This is exactly what double-blind tests are for, and being reminded of it may be the most useful yield of this whole project.
The story has a constructive epilogue, by the way, because the blind test showed not only what was missing but also where to go. When I had all the protection mechanisms removed again and ran the method pure, with a limit close to the value that Groh himself describes as a tuning rule in one of the patents (the correction as strong as possible while side effects just stop being audible), the first unambiguous music result appeared: eight out of eight on the string trio. The euphoria lasted one evening; then it turned out that half the effect was an artifact, an artificial stereo widening, because the correction was computed independently for left and right. After linking the channels, a small but reproducible remainder was left: seven out of eight at the previously named spot (guessing gets you there in only about three out of a hundred attempts), described as “the normal mix more composed, the Groh mix wider and more open,” measurable as 2.6 dB more side signal in the first ten-second passage. That is the honest state of affairs: the effect exists, it is pleasant, I prefer it in direct comparison. And it is small.
Someone Already Knew in 2013
And then the last piece fell into place. The research had surfaced a diploma thesis by Julian Klapp (FH Düsseldorf, 2013), who examined the IRT method back then with four professional recording engineers at Bauer Studios across eight recording situations, to this day the only external listening report on this invention. On first reading, still in the euphoria phase, I had mostly seen the confirmations: “more transparent,” “tidier,” “more stable to localize.” After the ABX series I read the thesis again, and now different passages jumped out at me. For the multi-miked single instruments, exactly the class of material I had tested, small-effect statements run through the interviews: differences “so small” that one expert, trying repeatedly, reaches no “consistent result”; another finds “the effort” of hearing a difference “too great.” In the polite language of qualitative interviews, those are the same five out of eight as mine.
The big effects in the thesis (“phasey,” “doubled,” all the way to “broken”) sit almost exclusively in a different class: vocals with a spaced main microphone system and a moving soloist, the classic main-plus-spot situations. Those I have not tested at all yet; I still lack the material, and that experiment is next, this time blind from the start.
Above all, though, the question this text began with now almost answers itself. Why did this patent never find a licensee? Presumably simply because on most material the audible effect is too small to carry a product. Add to that a problem of timing: as its target environment, the 2010 announcement explicitly names mixing consoles, a product world that was being displaced by the DAW in music production during exactly those years. And where, by all evidence, there would be the most to gain, with many open microphones on a stage, with summed lavaliers in theater, opera and television, we are mostly talking about live applications, and the method needs around 80 milliseconds of latency by design; the current implementation reports considerably more on top of that, owing to a safety margin for offline rendering. For post-production this matters less; for live use it is disqualifying, because there you count in single-digit milliseconds. So there are conceptual hurdles exactly where the benefit would sit.
What Remains
This is an interim report, not a final line. The vocal experiment is still to come, so are a few ideas about the latency, and I am not going to give up; I am having far too much fun. But it may well be that comb-filter-free summing remains what it probably always was: an invention in search of a problem. The history of technology is full of this type, and not the worst outcome would be for it to find its problem one day after all, perhaps in a niche that nobody in Munich was thinking of in 2009.
For myself, I take away two things, and they are bigger than the plugin. The first: the new agentic AI systems are, for someone like me, an empowerment I still lack the right word for. I am curious, widely read across fields, I see connections where there might be some, I have practical experience and a standard of quality, I can judge whether something is good, and I can draw the analytical consequences. What I always lacked was the capacity to actually execute every one of these ideas as a craft. That gap is closing right now. A method that sat in my head as a fascination for fifteen years, unreachable behind patents, inaccessible code and missing developer time, now stands, ten days later, as a tested, documented, honestly measured piece of software, including the finding that it can do less than hoped. That, too, is a result, and one that would simply have been unaffordable before.
The second is older than any AI: the ear is not a measuring instrument, it is an expectation amplifier. Between the noise test that electrified me and the five out of eight on the drums lies no technical insight, but one about myself. Double-blind tests are not a formality for scientists; they are the only tool that survives an inventor’s enthusiasm. Next time, I will start with them.
