Earshot’s involvement began within days. It was our first officially published investigation: Index, the independent investigative NGO leading the reconstruction of the case, brought us in to enhance and analyse the audio recorded in the five seconds before the shot. Fragments of speech were audible in the witness video, and their attribution had become a matter of national dispute. Our analysis identified two phrases, “Coupe” and “Pousse-toi”.

Earshot's enhancement isolating the two phrases

The second question was who had spoken them. Analysing the acoustic behaviour of both phrases, we found it consistent with voices outside the car. The voices, in other words, belonged to the police, not to Nahel.

The attribution analysis: locating the voices outside the car

These findings did not survive contact with the media unaltered. France 2 broadcast Earshot’s enhanced audio and told its viewers they were about to hear Nahel shouting “Pousse-toi” at the officers: an inversion of the analysis, recasting the evidence to fit the police narrative. The broadcaster later apologised. But the episode posed a question that has shadowed this case from the beginning: what happens to evidence once it enters the public domain, and who gets to decide what it means?

Three years on, that question is still open, most of all in the courts. On June 3, 2025, an investigative judge referred the officer to a jury trial for murder, an outcome almost without precedent in cases of French police violence. In March of this year, the Versailles Court of Appeal cancelled the murder qualification, retaining only a lesser charge. And on June 10, two weeks before the third anniversary of Nahel’s killing, France’s highest court broke that decision in turn, reopening the path to a murder trial. Nothing about this case is settled.

We spoke with Francesco Sebregondi, founder of Index, the NGO that has documented police violence in France since 2020. Francesco coordinated and directed the investigation into the killing of Nahel Merzouk from its first days: from the initial audio analysis published within a week of the shooting, through the year-long 3D reconstruction commissioned by Nahel’s family, to the evidence now sitting before the French courts.

Optional paid support: If you’d like to support our work, monthly subscriptions are available at $5.00. Earshot is the foremost independent nonprofit organisation producing sonic investigations with and on behalf of communities affected by corporate, state, and environmental violence. Your support enables us to conduct investigations, develop the field, and establish sonic evidence within accountability frameworks.

I. The first version

Earshot: The killing of Nahel Merzouk happened in broad daylight and was captured on video. And yet the images alone settled nothing. What did this case mean for Index as an organisation, and what does it say about the kind of work you see as your mandate?

Francesco Sebregondi: The first thing I want to say is that if we go back to the original unfolding of the story, on that very morning of June 27, 2023, the first account of events that emerged and circulated in the press appeared before the video did. The video only surfaced on social media in the afternoon. And that account was the relatively classic one for these police shootings at vehicles, for which we’ve had an epidemic in France since the passing of the 2017 law that introduced article L.435-1 of the Internal Security Code1. It was simply the police officer’s version: « the car drove directly at me, and I had no choice but to shoot at the windscreen in order to survive ». Only once the video emerged, and immediately contradicted that narrative, did the controversy really begin. And let’s remember how it began: with popular uprising. The very next day, an intense sequence of riots started across the entire country, triggered by the violence of the images themselves.

At Index, we had already been working on several cases of police shootings at vehicles, what French public discourse calls situations of refus d’obtempérer, refusal to comply. In 2022 there was a peak of 15 people killed in these situations. That peak is directly correlated to the 2017 law: researchers have since established a five-fold increase in people killed in these circumstances in the five years following the law, compared to the five years before2. It is a direct consequence of the loosening of the conditions under which police officers can open fire against moving vehicles.

The video went viral and immediately polarised French society. A lot of people saw an unjustified killing, under whatever legal framework you care to apply. But it was also very quickly reinterpreted by the far right as a right and just act: Nahel driving that car was a danger, and the victim would otherwise have been the officer, or others.

I can’t say we immediately perceived it this way, but this case is the ultimate stress test of the legal framework set up in 2017. For a self-defence argument to be legally valid, it needs to combine three conditions at once: necessity, proportionality, and immediacy. And it’s precisely on the temporal dimension that the law created a kind of dilation. It is formulated so that officers can open fire if the driver is susceptible of creating a danger. The immediacy question dissolves: whether the threat is immediate or could happen later in the future becomes a matter of the officer’s appreciation. And this situation encapsulated exactly that. Yes, the analysis of surveillance video could confirm that Nahel had been driving dangerously in the two-minute chase that preceded the shot: pedestrians were put at risk. But at the very moment he was shot and killed, he was simply restarting the car. There was no immediate threat to the officer himself; that is what the rest of our work established. The only threat was a potential one, in the future. That lethal force was used at that moment still poses a fundamental legal problem.

So it was clear this case had enormous ramifications, and it was important to analyse it and contribute. As an independent investigative NGO, beyond our focus on police violence, we understand our mandate as producing a counterpoint that is as factual, as documented, as rigorous as possible, from a civil society perspective, bringing into public discourse elements that can help society distinguish between the true and the false. That is a central political issue of our time.

II. The role of sound

Earshot: In the first days it was the sound, not the images, that dominated the public debate, as though the argument over what was said might settle what the footage already made visible. What made you decide to turn to audio analysis at that point?

Francesco Sebregondi: The fragments of conversation audible in the video were intensely debated. One can ask whether that debate wasn’t itself a kind of displacement, a media deflection from what was already there in the images. But since the controversy had emerged and was creating a heated debate, it seemed like an interesting way in. An audio enhancement could at least either clarify what was being said, or exclude some of the versions circulating at the time about what the police officers said before the fatal shot.

Earshot: What role did those fragments of speech play? And how has their significance changed as the case moved towards the courts? During the judicial investigation, Julien L., the second officer, was played the recording and admitted to hearing “balle sur la tête”.

Francesco Sebregondi: That admission concerns a different fragment from the one we identified ourselves. What was critical in the question of the transcription was whether we could establish, in addition to the homicide itself, some kind of intention or premeditation. If a threat of a bullet in the head was pronounced seconds before the materialisation of that threat, and this was later confirmed by one of the officers, that is an aggravating circumstance.

As for “Coupe” and “Pousse-toi”, the phrases from our own analysis, we were not able at the time to affirm in an absolute way that these were the exact terms pronounced. But we did challenge the alternative interpretations that were being circulated. For example, that it was an injunction to stop the car, which simply did not fit the syllables we could capture.

Earshot: One of the first things audible in the recording is the use of informal French, the officers addressing Nahel as tu. People across France heard that and attached their own experiences of the police to it. Was there a wider significance to how the officers were speaking to him?

Francesco Sebregondi: The question of tutoiement and vouvoiement also has to do with age, and here we are in a situation of adults facing children.

And given the wider context, a working-class banlieue near Paris, non-white passengers in the car, a two-minute chase just ended, nobody would have expected a courteous register in any case. But remember, we could not see the gestures in the original video, and this is a large part of what our investigation established by combining other angles. Their guns are drawn. They are striking inside the car: objects, or most likely people. The verbal aggression only reinforces the physical. The way they position themselves is already a threat.

Earshot: But three years on, the significance of those words has shifted.

Francesco Sebregondi: The legal defence of the police officer devolved in steps. The first thing he declared, and which was taken up in the press, was: “I was being run over, I had to shoot”. Then the video appeared and showed this was not true. So his defence became: “I was in danger, because by advancing, the car was potentially pushing me against the wall behind me”. At that stage, thatbecame the central argument to verify, and potentially invalidate, through analysis.

The question of the words was very important in the earlier framework, when the case had not yet entered a stricter judicial analysis. Because at the end of the day, he shot with a 9mm firearm at less than a metre’s distance. Whether or not he pronounced a death threat beforehand, it is still a killing that is at stake in the very act. Legally, everything turns on self-defence. Whether there is a circumstance that creates an exception within which that act can be legalised after the fact. And this is where the displacement towards the spatial configuration of the scene became absolutely central, and where the second layer of our work on the case unfolded.

III. Seeing beyond the frame

Earshot: The second, in-depth investigation went far beyond the audio: a full 3D reconstruction, cross-referencing the witness videos with the surveillance footage and the case file. How did that phase begin, and what did it establish?

Francesco Sebregondi: I can’t reveal information about our sources, but we somehow gained access to the full case file. It contained the surveillance video of the scene as well as other footage that had circulated far less, or that was found by the judicial investigation itself. So we had information at a much higher resolution with which to produce this investigation.

At the end of the day, it wasn’t particularly hard to contradict the officer’s argument. But it remains one of the clearest examples we use today of what this type of spatial, three-dimensional investigative analysis can do: it is a way to see beyond the frame of a video, to adopt a different perspective on an event. The image people associate most with our investigation, one that circulated widely on social media, is the moment we zoom into the frame, shift position, and reveal the space that had been flattened by the original video: the space between the police officer and the wall behind him. Combining all the different sources enabled us to prove, beyond any doubt, that the car was actually moving away from the wall when it restarted, and therefore posed no threat to the officer at that moment.

And what is interesting is the third evolution of the legal defence. Once the argument of I was in danger against the wall was dismissed, and this is what the lawyers who entered our investigation into evidence tell us, the only defence left is the one potentially offered by article L.435-1. So the officer’s defence shifted entirely towards that: it is no longer a question of me, of the danger for me. I did it because I needed to save other people’s lives, which were being threatened by the way Nahel was driving. By shooting, I saved lives. This is something the officer actually says at the end of his hearing. I don’t want to paraphrase too much, but it is words to the effect of: I regret that a seventeen-year-old boy was killed, but I am convinced that I saved lives that day.

This is where our work has little more to add. Once the facts are established, once the circumstances are clear, it becomes a purely legal and political question, the interpretation of those facts in legal terms. And this is where Nahel, as a case, presents the sharpest test yet of that 2017 law. Our hope, shared with the civil society groups mobilised against it, is that the ruling that will one day come in this case may also be the occasion of the law’s full abrogation.

IV. The trial that isn’t yet

Earshot: Three years on, Nahel’s story is playing out simultaneously in the courtroom and in the community. What has the case changed at the level of civil society?

Francesco Sebregondi: One important aspect to remember: within days of the killing, with the video circulating, a far-right figure, Jean Messiha, started a crowdfunding campaign for the police officer. It raised 1.6 million euros in a matter of weeks, money the officer’s family actually received3. The shooter was placed in preventive detention, then released, moved out to a nicer area, and came out a millionaire. So today, three years later, you will often hear it said, on social media or in informal discussion, that killing an Arab in France makes you a millionaire. That is how the message of this case has been received. And it is still fuelling enormous anger, beyond the strict legal framing of the judicial case: the fact that the killer of Nahel is both free and rich after what happened.

Earshot: And in the courtroom: Florian M. going to trial for murder would be an almost unprecedented outcome in a case of French police violence. Where do things actually stand?

Francesco Sebregondi: We are not yet sure he is going to trial. This is another important development. On June 3, 2025, something exceptional happened, an earthquake in the French legal system with regard to police violence: the investigative judge in charge of the case referred the police officer to a jury trial for murder. That decision was major, and was perceived by much of the community as a significant victory. A police officer facing a jury trial for murder is extremely rare in recent French history; in all of French history, in fact.

Then came the appeal. On March 5 of this year, the Court of Appeal in Versailles cancelled the murder qualification and retained only violences ayant entraîné la mort, violence leading to death. Honestly, even in French, the way it is formulated makes no sense. And an important consequence of this requalification is that the officer would no longer face a jury trial but a tribunal correctionnel, structured only by professional magistrates. No popular involvement, no jury. That led to another wave of mobilisation, marches. And to an appeal against the appeal, which, interestingly, every party subscribed to. Even the general prosecutor appealed the decision, as did the officer’s lawyer, who is still hoping for a full dismissal of the case.

So it went to the Cour de Cassation, the highest court in France, which ruled on June 10, just a few days ago. The Cour de Cassation broke the appeal court’s decision. All it can do is break a decision, not impose another one; the case returns to Versailles, which must now reformulate. Everyone reads the ruling as a rebuke of the requalification. Essentially: you have not justified your decision sufficiently. It seems very unlikely, from the readings I’ve had from different lawyers, that the Court of Appeal could reissue the same requalification, because there is little chance they can justify it more strongly than the first time. So the most likely outcome, perhaps in another six months, is a confirmation that Florian M. will face a jury trial for murder. Likely, but not yet guaranteed.

Earshot: And all of this is unfolding on a particular political calendar.

Francesco Sebregondi: We are entering an electoral year in which there is a strong plausibility of the election of a far-right president. And the far right has erected the Nahel case as a symbol of the choice French society has to make. This framing is taken up directly by the officer’s lawyer, who says: you have to choose between a society of Nahels and a society of police officers who represent order. So the long extension of the judicial process, potentially beyond the electoral deadline, is in itself a threat to the possibility of the process being led to its end. The judiciary is independent in principle, but should a far-right president be elected, we would be entering uncharted territory with regard to the application of the rule of law.

And this is simply a constant in the judicialisation of police violence in France, especially police killings: cases extend over five, six, ten years… This extension of the judicial process is part of the mechanism of impunity and protection of police officers within the French legal system.

V. The distance between law and justice

Earshot: This case shows everything you’re able to do: tell the story, mobilise social media, contradict the news, force erroneous reporting into apology, intervene at every level. But it feels like your technical expertise then meets a « final boss »: legal technicality. A shadowy process in which everything you can establish, all the facts, could simply evaporate. Is that something you see often in your work, or are there cases where you crack through the legal gymnastics?

Francesco Sebregondi: That’s a great question. Based on our experience (six years, and we just counted, twenty-three cases of police violence within the French legal system) the lesson we draw is that we cannot let the courts be that« final boss », that ultimate interlocutor. We should understand our work, rather, as contributing to disclose and expose the inherent limits of the legal system to handle these cases: to actually produce justice. We’re not expecting much from the judicial system. We fight those legal battles alongside families of victims, and we’ve even won some of them, though you have to constantly question which criteria you use to define victory when it comes to a legal process. But it’s really about the fight itself, and about shifting the limits, the conditions, the borders of where the battle happens.

A case like Nahel’s shows how that works. And here I want to make sure the full picture includes the thousands of young people that took part in the riots in the first days; the families of victims and the committees around them that have been at the forefront of the mobilisation around police violence and police racism in France; the allied media. This popular and public force ended up completely overwhelming the set of limits the state wants such a process to remain within. The battle is still being fought, but a lot of the victories have already been won. The arena within which this debate can be held in French society has already been reconfigured.

Earshot: There’s a way in which the analysis exists to expose the facts on the ground, of course, but also to measure the distance between the law and justice. Like looking between the wall and the police officer: it’s that measurement, that space, that gap. The facts of that morning are no longer in dispute. What remains in dispute is what the law will do with them. This is precisely the threshold at which we, Earshot and Index, operate: investigations that make the same distance measurable, between what power says happened and what happened, between law and justice. The measurement does not close the gap, but it makes it harder to deny, and more costly to maintain.

Francesco Sebregondi: Exactly: it’s exposing those gaps. If one of the mechanisms for maintaining the impunity of police forces, and of state violence in general, is to legalise it after the fact, then our work is to make that legalisation costly. Not to prevent it completely, we know that’s not going to happen, but to make it as costly as possible: to expose the contradictions of a system that, at some point, can force the state to reconfigure itself.

The full reconstruction: Index's investigation into the killing of Nahel Merzouk, in French with French subtitles (16 min)

Across our cases we have come to see a problem in the field of audio authentication: the prevailing practice is to rely on detection software alone. A clip goes in, a verdict comes out – real or fake – and that verdict is treated as the answer. Our own work, which combines critical listening with detection software, is increasingly weighed against the outputs of these tools, as if the software’s verdict and a careful acoustic reading were equivalent forms of evidence. This assumption obscures what authentication requires.

In this month’s newsletter, we share reflections on a few of the cases where we set out to determine whether a recording was genuine or AI-generated, and on the methodology we have been developing for audio authentication of deepfakes: one that pairs critical listening with detection software, attending to the sounds in and around the voice.

Optional paid support: If you’d like to support our work, monthly subscriptions are available at $5.00. Earshot is the foremost independent nonprofit organisation producing sonic investigations with and on behalf of communities affected by corporate, state, and environmental violence. Your support enables us to conduct investigations, develop the field, and establish sonic evidence within accountability frameworks.

Methodology: Critical listening and detection software

The AI models that produce synthetic speech, and the detection tools built to catch them, are trained almost entirely on the voice: its frequencies, rhythm, timbre, the fine patterns that make a person sound like themselves. The models can only reproduce, and the detection tools can only listen for, what they have been trained on. The voice, though, is but one part of what a microphone actually records.

Around the voice sits everything else: breaths and hesitations, the incidental sounds on a recording, the resonance of the space, the strain of the device itself. An AI generation model trained on the voice alone has no room to resonate, no environment to draw incidental sound from, no real air to move across a microphone diaphragm. These surrounding sounds are not noise to be filtered away; they are, we have found, often more reliable evidence of a recording’s authenticity than the voice. It is precisely this relational web of sound, rather than the voice, that current models fail toreproduce and that binary detection tools are not built to hear.

Our methodology is critical listening: close attention to the audio artefacts surrounding the voicethat together form the coherent acoustic web of a genuine recording.

Where it is available, we use detection software as a supplement: a fast, quantified second opinion that examines the voice. But the software is not the work. As the cases below show, some of the most decisive evidence in our investigations is rarely linguistic, and almost never come from the voice itself.

What follows is not a full account of any single investigation. We have isolated each artefact in turn and paired it with the case in which it proved most decisive, to show what each one does and why it matters.

Breaths

Breaths are typically heard lasting under a second and are characterised by their broadband frequency content where no single frequency dominates, but rather an airy mixture of all frequencies is heard across the audible spectrum (0–20kHz). Spectrographs allow us to visualise this behaviour. A spectrograph is a visual representation of the frequencies present in an audio recording, with frequency mapped on the y-axis and time on the x-axis. The bright orange colour represents the intensity of each frequency at a particular moment in time. When we examine a spectrograph of a natural breath, the broadband frequency content appears as a diffuse cloud – the product of chaotic air movement around a microphone, causing uncontrolled vibration across all frequencies (Figure 1).

In November 2024, three audio clips were shared by members of the ruling political alliance ‘Mahayuti’ in Maharashtra, India. These clips contained conversations between the leaders and officers from the opposition party ‘Mahavikas Aghadi’ purportedly promising financial favours during an election in the Maharashtra Assembly1. While the voices in these recordings were remarkably close to the people they were targeting, our analysis confirmed they were AI-generated by identifying a strange pattern in the sounds of breaths between sentences.

An AI-generated breath, with the rhythmic, tremolo-like undulation in amplitude that gave the Maharashtra recordings away.

Unlike the natural breaths we were accustomed to hearing and seeing in our spectrographic analysis, the breaths in the Maharashtra investigation were showing a different pattern: a repeating pattern of pronounced frequencies appearing at intervals of a few milliseconds, visible as vertical bands across the entire breath (Figure 2).

These rapid fluctuations in volume resemble the tremolo effect applied to a guitar – a rhythmic sound undulating in amplitude, not commonly found in natural speech. A seemingly innocent exhalation of air between alleged admissions of financial fraud holds the key to the larger corruption at play: the defamation of the opposition party by those in power. Artificial intelligence, it turns out, is still learning to breathe.

A natural breath, for comparison: broadband and chaotic, the sound of real air moving across a microphone.

Room resonance

Acoustic artefacts are also produced by the unique architecture of the space where recordings are made. Every enclosed space has a ‘modal resonance,’ or a set of frequencies at which sound waves reflect and reinforce each other. This is determined by the room’s dimensions, shape, and construction materials. When a sound containing those specific frequencies is introduced into the room (e.g., by a loudspeaker or a crowd cheering), the room is acoustically “excited,” and those frequencies will ring out longer and sound louder than others.

In March 2026, we analysed a recording in which a top aide to Ali Khamenei is heard addressing IRGC commanders, in the period following Khamenei’s death in a US-Israeli strike. The aide recounts a separate strike on the home of Khamenei’s son Mojtaba – aimed at him in anticipation of his succession – and how Mojtaba had survived. With Mojtaba subsequently elected as Iran’s new supreme leader, authenticating this recording took on significant weight. With no public appearance to corroborate his survival, this audio was the regime’s principal channel for asserting it.

Enhanced recording showing the room resonance activated by the aide’s voice and by the response of the audience, probably IRGC commanders, at the end of the recording.

Our analysis found that the recording was likely authentic. We arrived at this conclusion by analysing resonating frequencies in the aide’s voice, heard on amplified speakers in the room, and in noises from an audience, probably IRGC commanders, in the background of the recording. The aide’s amplified voice consistently activated the specific resonant frequencies of the room, most prominently at 200 Hz, 400 Hz, and 700 Hz. Crucially, the audience’s noise present at the end of the recording activated the identical set of resonant frequencies, confirming that both the aide and the audience were in the same acoustic space (Figure 3).

Real-world recordings are rarely without room resonances and AI generation models do not make the distinction between the frequencies of the voice and the resonance caused by the surrounding environment. Synthetically generating both speech and crowd noises with the same resonant frequency would require the model to train on recordings that already contain this same resonance, or skilled manual postproduction. Though not impossible, such a construction is unlikely.

Microphone distortions

Acoustic artefacts are produced at multiple scales – from the resonance of a large hall to fleeting sounds in and around the small components within a microphone. Microphone distortions are typically caused by physical interactions with the recording device: either through physical contact, or through quick changes in air pressure surrounding the device caused by speech. This latter form of distortion occurs when a speaker exhales strongly and directly onto the microphone, or even when pronouncing with emphasis consonants such as ‘p’ or ‘t’ and the breathier ‘f’, ‘h’ and ‘sh’. These interactions cause the diaphragm to ‘pop’ under the force of the air. While common in real-world recordings, replicating this artefact in AI-generated speech would require either deliberate insertion, or for the algorithm itself to be scripted to include the sound of a ‘pop’ around particular consonants.

In May 2026, we analysed a leaked audio from Brazil that allegedly captures the voice of Flávio Bolsonaro, a right-wing candidate in upcoming general elections. The audio was purportedly sent to Daniel Vorcaro, a banker under investigation for financial fraud. If authentic, the recording would show a previously unconfirmed connection between the banker and the Bolsonaro family.2

Our analysis identified multiple distortions caused by the speaker exhaling directly onto the microphone, but the distortions did not stand alone. The recording also contained a further patternwe look for routinely in our work: incidental sounds, such as vocal hesitations (‘mm’ and ‘eeh’), tongue clicks, swallowing, throat clearing.

A vocal hesitation (’eeh’) followed by a microphone distortion caused by an exhalation onto the microphone. From the recording allegedly featuring Flávio Bolsonaro.

While AI models may be able to generate these sounds, they currently require such sounds to be inserted manually into a text prompt:

‘I am sitting in a room [sigh], different from [hmm] the one you are sitting in now [clears throat].’ 34

In either case, the staging is conscious where natural speech is instinctive, and the density required to match a real recording – hundreds of small sounds at unpredictable intervals – quickly becomes impractical to sustain.

The Bolsonaro recording was thick with these incidental sounds. Together with the microphone distortions, these patterns formed a rich acoustic backdrop consistent with authentic speech (Figure 4).

These same methods were recently applied in our investigation into the Hondurasgate scandal, in which leaked audio revealed US interventions aimed at sidelining leftist governments in Central and south America.5

The software in detail: Resemblyzer

As the methodology section established, detection software supplements our critical listening when it helps. Here we look at one such tool in detail – Resemblyzer from Resemble AI – to show how these systems work and where they fall short.

We focus on Resemblyzer as it is open-source and narrower in scope, but there are other tools available out there, such as Phonexia (used widely in forensic voice biometrics) and Azure Speech, Microsoft’s speech analytics platform. These differ in scale and approach, but the underlying logic is the same: feed the software a sample, get back a number or a category.

Resemblyzer takes a recording and reduces all the patterns of a voice that distinguish one speaker from another into a single numerical value, a kind of vocal fingerprint. That fingerprint can then be compared against another recording of the same supposed speaker, and the comparison returns a score between 0 and 1: the closer to 1, the more likely the two voices belong to the same person; the closer to 0, the more likely they do not, and the more reason to suspect synthetic generation.

What Resemblyzer attempts is essentially the automation of phonetic analysis: pattern-matching the features of speech that distinguish one speaker from another. Used carefully, the score is useful as a cross-check: something to read against what we are hearing. But two limitations sit inside the number, and both become visible only once you understand how the score is produced.

The first is that the scale has no fixed threshold. There is no point on the 0-to-1 range at which a recording becomes definitively fake; a score of 0.6 is not a verdict, only a prompt for closer attention. Where the line falls depends on the voice, the recording conditions, and the quality of the reference material. This shifts as generative models improve, and synthetic voices score higher. The number narrows the question; it does not answer it.

The second limitation lies not just with Resemblyzer but with the field as a whole. Tools like Resemblyzer are trained predominantly on a handful of well-resourced languages, English above all. Most of the recordings that reach us are in Spanish, Portuguese, Farsi, Hindi, and Arabic amongst others, often in regional accents the model has barely encountered. So, the tool is least reliable in the languages our cases most often come in.

These two limitations point to a distinction at the heart of our work. Resemblyzer, like the broader practice of phonetic analysis, asks whether speech is consistent with a given speaker. This is the work of linguists and phoneticians, and they would not rely on a single similarity score to answer it. Their professional protocols call for multiple methods, careful comparison against reference materials, and judgement that cannot be reduced to a number. Yet in media and elsewhere, organisations increasingly do rely on automated tools alone, treating the score as the verdict it cannot honestly be.

Even at its best, this is speech analysis. It asks whether a voice is a specific person’s. Audio authentication asks a different question: whether a recording is what it claims to be. To answer it, you have to attend to everything the microphone captured beyond the voice, the world the voice inhabits. As the cases above show, it is that world – sometimes alone, sometimes in combination with the voice itself – that reveals whether a recording is authentic.

Listening for coherence

Across these cases, a single pattern holds. The evidence that settled each one rarely came from the voice alone. It came from around it: the breath too regular to be real, the room that resonated exactly as an actual room does, the microphone clipping under the force of an exhalation. AI-generated speech, for now, does not usually fail because the voice is wrong. It fails because the voice does not belong to the world it claims to inhabit.

This is the part of authentication that is hardest to automate, and the part most often missing from the conversation about it. The prevailing tools treat the problem as closed and binary, and they are built for a handful of well-resourced languages. As speech-generation tools flatten regional difference in the name of fluency, the gap widens precisely where the stakes tend to be highest. A better score will not close it. What will is a wider practice: methods that can be examined and contested, documentation of how conclusions are reached, careful attention to the acoustic conditions of a recording beyond the voice, and collaboration with linguists who know the languages and places these recordings come from.

Synthetic recordings will keep coming, and they will keep getting better, but none of that will replace the need for the practice of critical listening. Placing two modes of digital analysis in contention with one another is not the answer – both can only describe the slice of a recording that machines can model. What reveals whether a recording is authentic sits in the excess: the room, the breath, the microphone, everything the voice happens to inhabit.

Authentication, in the end, is the work of listening for coherence: whether a recorded voice belongs to its room, its moment, its claimed world.

Optional paid support: If you'd like to support our work, monthly subscriptions are available at $5.00. Earshot is the foremost independent nonprofit organization producing sonic investigations with and on behalf of communities affected by corporate, state, and environmental violence. Your support enables us to conduct investigations, develop the field, and establish sonic evidence within accountability frameworks.

5

Drop Site News

Hondurasgate: Key Leaked Audio Files, Revealing U.S. Intervention in Honduras, Found Authentic "With Moderate Confidence"

President Donald Trump’s brazen intervention in November 2025 in the Honduran presidential election to openly call for a vote for the National Party’s Nasry Asfura came as a surp…

Read more

3 months ago · 484 likes · 18 comments · José Luis Granados Ceja

(Accessed May 2026)

El-Hitami’s article moves between three cases that, at the time, captured where our work was headed. The killing of Nahel Merzouk in Nanterre, which we worked on in collaboration with INDEX, and where a five-second audio clip held the answer to a question the footage couldn’t resolve, anchors the piece. The October 2023 strike in southern Lebanon that killed Reuters journalist Issam Abdallah sits alongside it, with the drone analysis we carried out for Amnesty International and Human Rights Watch. And a third thread, woven through the piece, captures the more informal side of our work: our investigation into the killing of Hind Rajab and her family in Gaza, whose findings we shared with the Washington Post in response to their questions, and an exchange with Liminal in Italy, walking them through our methodology for earwitness interviews so they could apply it in their own investigations.

Reading it back now, what stands out is how much of our approach was already in place: the insistence on probabilities rather than certainties, the commitment to transparent methodology, and the conviction that acoustic evidence deserves the same forensic seriousness as the visual. A lot has happened in the eighteen months since, but the foundations El-Hitami captured are still the ones we’re building on.

Optional paid support: If you’d like to support our work, monthly subscriptions are available at $5.00. Earshot is the foremost independent nonprofit organisation producing sonic investigations with and on behalf of communities affected by corporate, state, and environmental violence. Your support enables us to conduct investigations, develop the field, and establish sonic evidence within accountability frameworks.

The collection of evidence

Earshot uses acoustic evidence to solve crimes: the sound of a gunshot, the high-pitched humming of drones. The organisation aims to solve cases involving human rights violations where the state itself is the perpetrator.

One sunny morning in June 2023, Nahel Merzouk is shot dead. During rush hour in the Parisian suburb of Nanterre, a police officer points his gun at the 17-year-old who is travelling in a yellow Mercedes with two friends, and kills him at point-blank range.

Passers-by film the incident. One video shows two police officers leaning through the driver’s window. When Merzouk suddenly accelerates, a shot rings out. The car travels a little further, before crashing into an obstacle and coming to a halt. At quarter past nine, Merzouk is dead.

The fatal shooting of the French teenager of Moroccan-Algerian descent sparked fierce protests in France. Many saw it as a symptom of a racist system that treats young people from migrant backgrounds with excessive force. The United Nations has accused the French police of structural racism.1

It is beyond doubt that a police officer killed an unarmed minor. But the footage does not show why the officer fired. Was it self-defence, as the officer claimed afterwards? Or was it murder, an example of racist police violence? The answer may lie in a five-second audio recording of the exchange between the officer and the victim immediately prior to the incident.

“The entire debate revolved around this section here, which remains incomprehensible”, says Lawrence Abu Hamdan, pointing to a few blurry orange spots on his laptop screen.

The spots resemble car headlights in the dark. These spots form part of a spectrogram, a visual representation of sound frequencies. Where the spots become denser, someone can be heard speaking. But who? And what? Due to the poor audio quality of the mobile phone videos, this was initially impossible to discern, says Abu Hamdan.

Lawrence Abu Hamdan, 39, describes himself as a “private ear”, a play on the English term “private eye” for a private detective. This is because, in his investigations, he pursues acoustic evidence. About a year ago, he founded the not-for-profit organisation Earshot. The killing of Nahel Merzouk was the first case for the three-person team. Abu Hamdan and his colleagues used special software to digitally enhance and analyse the crucial seconds of the recording. What they discovered could help solve the case.

In early April, the team meets in London. Abu Hamdan is wearing a brown-and-blue lumberjack shirt and thick-rimmed glasses. Visually, little may remain of the teenager who played in a punk band, but the passion for sound and the political awareness from that time remain with him to this day.

Some of Earshot’s work is based on the methods of forensic linguistics, a discipline that deals with language in the context of criminal investigations. Yet, while forensic linguists typically work with state law enforcement agencies, Earshot is dedicated to investigating human rights violations perpetrated by the state.

“In human rights work, acoustic evidence is still perceived as inferior to visual evidence”, says Abu Hamdan. There is simply no expertise available to analyse the acoustic material. Without Earshot this gap would remain. “There is no one else doing this work”.

For more than 15 years, the British-Lebanese researcher and artist has been engaged in acoustic research. He completed a PhD on the subject, has collaborated with organisations such as Amnesty International, and has published research in international media. He has also presented his findings in art exhibitions, for which he received the Turner Prize in 2019.

Abu Hamdan and his team usually work from an office in south-east London. As the power lines are currently being repaired there, Abu Hamdan is sitting in his sister’s flat today. Together with his colleague Fabio Claudio Cervi, he has set up in the living room at a large wooden table. They have set up their so-called ‘supercomputer’ and a microphone; on the table lie sound cards that can be connected to the PC, a pink sketchbook, headphones and memory cards, with the last square of a bar of chocolate in between. Cervi is currently working on a long-term project for Earshot: a video-game simulation about noise pollution. While he is programming, he and Abu Hamdan answer questions from a journalist from the Washington Post, on whose behalf they have carried out an acoustic analysis.

“French sometimes listened to a single syllable for three days” — Abu Hamdan

Abu Hamdan entered the world of audio investigations in 2009 following a meeting with forensic linguist Peter French, who had been involved in thousands of police investigations. In 1984, police interviews in the UK had to be recorded for the first time. Suddenly, courtrooms were debating exactly what a defendant had meant, rather than simply accepting the written police record. “French would sometimes listen to a single syllable for three days,” says Abu Hamdan. “His attention to linguistic detail was groundbreaking.”

French investigated, for example, whether defendants had said different things in court than in police interviews, or whether their strong dialects had caused misunderstandings. Later, he and his colleagues were tasked with investigating beyond literal statements. They listened to background noise on emergency calls, for instance, to gather information about the crime scene. “In this way, the linguists broadened their auditory horizons, analysing not only human voices, but also the voices of bodies, architecture, ammunition and infrastructure,” wrote Abu Hamdan in his doctoral thesis.

Earshot’s work also goes beyond the analysis of language. Among other things, the team has developed a tool for analysing gunshots recorded by mobile phone cameras, for example. The “Supersonic Compass” analyses two things: the shockwave of a projectile flying through the air, and the explosion at the muzzle of the rifle when the trigger is pulled.

“Our tool simulates the movement of the projectile and, using the position of the recording camera, reveals the shooter’s location,” explains Cervi.

This allows the statements of eyewitnesses and earwitnesses to be corroborated. Furthermore, Abu Hamdan, Cervi and their colleague Caline Matar regularly verify the authenticity of audio recordings. To do this, they listen, for example, for electromagnetic interference that can be heard in the background of recordings.

“We work with the excess: we listen to the noises that the brain normally filters out.”

In a corruption case, Earshot was able to establish that leaked recordings of government officials were authentic.

“The background noise remained consistent throughout the entire recording,” explains Abu Hamdan. No section had been removed or added afterwards.

A notification pops up on Cervi’s screen: the Italian organisation Liminal is calling because they want to find out what drones and other aircraft migrants encountered while crossing the Mediterranean. Abu Hamdan explains how to properly interview earwitnesses: don’t assume too much, but let them recall things themselves. Offer various sounds for comparison.

Analysing drone sounds is part of everyday life at Earshot. In war zones such as Ukraine or Gaza, the high-pitched hum of drone engines is omnipresent. At the end of 2023, Earshot analysed the sounds of an Israeli drone in collaboration with Amnesty International and Human Rights Watch. The drone could be heard in a video recording made by journalists who were struck by an Israeli tank shell in southern Lebanon. Reuters reporter Issam Abdallah was killed, and six other journalists were injured.

Abu Hamdan plays the drone noise on his PC: a piercing whirring, a harbinger of death and the acoustic epitome of automated warfare.

“The waveform of the frequency tells us something about the movement of this drone,” explains Abu Hamdan.

The sound gets lower as the drone moves away. It becomes higher-pitched and louder as it approaches. Because the engine is located at the rear, the drone is loudest at the moment it turns away and flies off. “Based on the volume and frequency, we realised that the drone had been circling above the journalists for 23 minutes before the attack.” During this time, the attackers were able to gather enough information to determine that their targets were not military, but rather the press.

The cases that Earshot takes on could be pursued more easily by the relevant states. Law enforcement agencies would have the opportunity to access, for example, the video recorded by the drone’s camera or the bodycams of police officers who killed or injured people. Earshot’s work, however, begins where the state drops, manipulates or drags out investigations — or where there is concern that this might happen.

Abu Hamdan plays the drone noise on his PC: a piercing whine/whirring, a harbinger of death and the acoustic epitome of automated warfare.

As in the case of Nahel Merzouk, whom a police officer in Paris killed with a bullet to the chest. Although the investigation into the officer is still ongoing, he has been at large since November 2023. Statistics show that cases against police officers are dropped far more often than those against other suspects. It is cases like these that Abu Hamdan and his colleagues listen to closely.

“Even if our evidence doesn’t make it to the courtroom, we can at least exert pressure by standing up for the results of our investigations and bringing them to the public’s attention.”

Once the Earshot team had processed the audio track from the videos of the incident, it became clear what was said before the shot was fired: “Coupe”, the French command to switch off the engine. And “Pousse-toi”, colloquial for: “Move out of the way”. The context and slang might suggest that the teenager Merzouk shouted “Move out of the way!” before stepping on the accelerator to flee, which posed a danger from the police officer’s perspective. This was initially claimed by the French media as well.

Earshot, however, discovered that the two statements must have come from the same voice. A second mobile phone video from the scene of the incident shows that it was the police officer who said “Coupe”. The fact that the frequencies of both statements match makes it clear that he also said “Pousse-toi”.

“If the voice came from inside the car, was reflected off the wall and then picked up by the mobile phone’s microphone, it would look completely different on the spectrum from the voice outside the car.”

Earshot does not claim that its analyses are exhaustive or represent absolute truth. “We only ever talk about probabilities,” says Abu Hamdan. “And we make it transparent how we arrive at our conclusions.” An investigation like the one into Merzouk cannot solve a case, but it can raise crucial questions.

For example: Why did the police officer say “Get out of the way” to a young man whom he had just ordered to switch off the engine, whilst pointing a gun at him? Abu Hamdan has a theory: “If the police officer directed the command not at Merzouk but at his colleague — ‘Get out of the way, because I’m about to shoot’ — then Nahel understood at that moment what was going to happen and drove off to save his life.”

Merzouk is said to have told his passenger: “He’s crazy, he shot at me.” Then he died.

📄 Full investigation: For complete technical analysis and sound propagation modeling, read our assessment here: Police Brutality in Nanterre, Paris

At the epicentre of the second strike, Reuters photojournalist Hatem Omar was filming. When the blast hit, his camera was buried in the rubble, but it kept recording. Omar survived, badly wounded. Though the image shows almost nothing, its soundtrack became central to this investigation.

Since the beginning of the genocide in Gaza in October 2023, the Israeli army has been systematically targeting journalists and press infrastructure across Gaza, the West Bank, and Lebanon. The Committee to Protect Journalists recorded 86 journalists killed by Israeli fire in 2025 alone1. Our work has documented attacks on journalists across multiple cases, including the October 13, 2023 strike in South Lebanon that killed Reuters journalist Issam Abdallah and injured six of his colleagues, and five further cases documented in collaboration with Forbidden Stories2. In each, the same pattern: the Israeli army knowingly conflating press with military targets.

When cameras are targeted — buried in rubble, choked by smoke, their operators killed or forced to flee — the image fails, but sound persists. The soundtrack captures everything the image can no longer reach: the direction of a blast, the weapon used, the distance of a shooter. By listening closely to the sounds of cameras under fire, we can reconstruct what journalists could no longer survive to show and, in doing so, restore the conditions of witnessing under fire. What follows is one such reconstruction.

Optional paid support: If you’d like to support our work, monthly subscriptions are available at $5.00. Earshot is the foremost independent nonprofit organisation producing sonic investigations with and on behalf of communities affected by corporate, state, and environmental violence. Your support enables us to conduct investigations, develop the field, and establish sonic evidence within accountability frameworks.

Double-tap strike on Nasser Hospital, Khan Yunis, Gaza

Following the 10:17 am strike, we were approached to verify a claim circulating in early media coverage that the weapon used may have been a Lahat missile — a guided munition capable of being fired from tanks, drones, or helicopters. Our audio ballistic analysis of five videos documenting this strike rules out that either of the two projectiles fired at Nasser Hospital was a Lahat missile, and identifies both as tank shells. The first of the two projectiles was likely an M329 set to anti-personnel mode, a weapon whose design and deployment we examine in detail below. Our analysis further establishes the likely distance from which the first projectile was fired with the direction from which it approached.

I. The weapon used

By comparing the acoustic signature of the recordings against that of a Lahat missile, we were able to rule it out and identify what was likely used instead.

If the weapon used was a Lahat guided missile, one would typically hear a distinctive whistle as it descends: the sound of the jet propulsion carrying it to its target. The spectrograph below shows what that sounds like: the prominent whistling of the jet engine preceding the detonation of a Lahat missile.

In all five videos, no such sound is present. What we hear instead is something we first encountered in our investigation into the strike on journalists in South Lebanon on October 13, 2023: six distinct detonations occurring within approximately 300 milliseconds of the initial blast. This is the acoustic signature of the M329 set to anti-personnel mode, confirmed in the South Lebanon case by visual evidence of the tank and forensic analysis of fragments recovered from the scene3. We subsequently identified the same signature in an attack on civilians gathering to collect humanitarian aid at Kuwait Roundabout in Gaza on February 6, 2024.

The Nasser Hospital recordings show the same pattern. It is therefore highly likely that the first of the two projectiles that struck the hospital on August 25, 2025, was an M329 set to anti-personnel mode.

This acoustic finding is corroborated by physical evidence recovered at the scene. Reuters obtained photographs of metal fragments found at Nasser Hospital, taken by a doctor present that day. Five munition experts who reviewed the fragments independently identified them as tail fins of Israeli-made 120mm tank rounds4. This is consistent with our acoustic identification of the M329.

The M329 is a 120mm multi-purpose tank round developed by Elbit Systems. It contains six internal submunitions, each lined with approximately 4,000 tungsten cubes measuring 2mm. In anti-personnel mode, these submunitions are released unguided in mid-air and detonate sequentially across the target area, each spraying tungsten fragments in all directions.

The Israeli army chose a weapon designed to disperse lethal fragments across the widest possible area rather than a guided missile aimed at a single target. That choice is not incidental. In a double-tap strike, the goal of the second projectile is to kill and injure as many as possible among those who respond to the first. The M329 set to anti-personnel mode is precisely the weapon for that purpose — making this a textbook double-tap strike. Its acoustic signature does not only identify what was fired. The selection of this weapon confirms the intention to commit this illegal act.

II. Where it was fired from and how far away

Knowing what struck Nasser Hospital, the recordings allow us to go further: to establish the likely location of the tank from which the projectile was fired.

When a tank fires, it produces two distinct sounds: the detonation of the shell at its target, and the muzzle blast, which is the explosive release of gas from the barrel as the shell is fired. Although both sounds originate from the same firing event, they travel at different speeds. The shell travels at supersonic velocity; the muzzle blast travels at the speed of sound. This means that at the location of impact, the muzzle blast always arrives after the detonation. The gap between the two — measured at the point of impact — represents the time it took the muzzle blast to travel from the tank to the camera. Combined with the known velocity of the shell, this interval allows us to calculate the distance of the tank that fired.

In three of the five videos analysed, each recorded from a distinct location around the hospital, we detected a high-intensity impulse sound in the low-frequencies (0-500 Hz) arriving approximately 3.5 seconds after the first blast. That this sound appears across all 3 recordings simultaneously, each taken from a different position, makes it highly unlikely to be an artefact of any single recording. Its low-frequency profile and consistent timing across locations point to a muzzle blast, most likely from the tank that fired the shell.

II.I. Distance

On the morning of August 25, 2025, the temperature in Khan Yunis was approximately 31 degrees Celsius. At that temperature, the speed of sound is 349 metres per second.

The M329 tank shell has a published muzzle velocity of 900 metres per second. A shell does not maintain this speed over distance. Air resistance causes it to decelerate continuously from the moment it leaves the barrel. Applying a standard ballistic drag model to the M329’s specifications (a 17-kilogram projectile, 120-millimetre calibre, fired at 900 metres per second into air at 31 degrees Celsius) the shell’s average velocity over approximately 2 kilometres is around 827 metres per second, arriving at the target at roughly 761 metres per second.

The distance to the firing position is calculated from the muzzle blast delay: the gap between the shell’s detonation and the arrival of the muzzle blast at the camera. Because the shell travels faster than sound, it always arrives first. The delay represents the difference between the time the sound took to travel from the tank gun to the camera, and the time the shell took to cover the same distance. Using the 3.56-second muzzle blast delay recorded at Video 2, positioned at the epicentre of the strike, and accounting for the shell’s deceleration over that distance, our calculation places the tank that fired at approximately 2.0 to 2.4 kilometres from Nasser Hospital.

Avihai Stollar, an arms specialist with Breaking The Silence who served in an IDF infantry unit, told the Washington Post that M329 rounds are accurate up to approximately 3.5 kilometres. At a distance of 2 to 2.5 kilometres, he said, tank crews would have been able to see people gathered at the scene. Whoever ordered the fire “could definitely see what is happening and could see rescue workers moving up and down the stairs of the building.”5

II.II. Trajectory

As three of the recordings contain a detectable muzzle blast, we were able to go a step further than calculating distance alone and estimate the direction from which the shot was fired. To do this we applied a technique known as multilateration. Multilateration, akin to triangulation, entails measuring the time difference between the arrival of a sound at two or more recording positions and converting those time differences into a direction. The variance in time it took the muzzle blast to reach each of the three cameras allowed us to determine the direction of the firing position relative to Nasser Hospital.

Sound travels at a consistent speed in a given environment. When the same sound arrives at multiple recording points at slightly different times, those differences reveal the direction it came from: it reaches the points closest to its source first, and those furthest away last.

The muzzle blast arrived at Video 1, 3.4 seconds after impact. It arrived at Video 2, positioned at the epicentre of the strike, 3.56 seconds after impact. And at Video 3, 3.58 seconds after impact. Because the three cameras were recording independently, their clocks were not synchronised to a common reference. To make the arrival times comparable, a correction was applied using each camera’s GPS-derived distance from the impact point: 49.43 metres for Video 1, 60.39 metres for Video 3, and zero for Video 2, which was positioned at the epicentre of the strike. Factoring in these corrections, the muzzle blast reached Video 1 first, then Video 3 approximately 149 milliseconds later, and Video 2 a further 153 milliseconds after that.

Video 2 is the westernmost of the three cameras; Videos 1 and 3 are positioned to its east. The muzzle blast reached the eastern cameras first and the western camera last. That arrival order, combined with the known positions of the three cameras, places the origin of the sound to the north-east of the hospital.

Our acoustic analysis accounting for the shell’s deceleration and the margin of uncertainty in the muzzle blast delay measurement, places the firing position between 2.0 and 2.4 kilometres north-east of the hospital. The Washington Post independently identified an active IDF patrol base approximately 2.5 kilometres north-east of Nasser Hospital. Their analysis of satellite imagery provided by Planet Labs showed Merkava tanks and armoured personnel carriers visible over a 12-day period spanning the date of the strike6. When we calculated the distance to the coordinates provided by the Washington Post, the base extends approximately 150 kilometres in depth along the north-east axis, with its far edge sitting at 2.4 kilometres from the hospital — meaning it falls entirely within our estimated range. The distance and direction established by our analysis are therefore consistent with that base, and it is highly likely that the strike on Nasser Hospital was conducted from this location.

III. Doctors’ testimonies on tungsten submunition injuries

The testimonies that follow were gathered as part of a study on weapons similar to the M329 that contain these tungsten shrapnel conducted with the support of Forensic Architecture’s researchers in Gaza, in which six international doctors who worked in Gaza in 2024 described the injuries they treated. The tungsten cubes they recovered from patients — documented by The Guardian7 — are small enough to be nearly invisible on entry.

Dr Nabeel Rana, a surgeon who worked at Al-Aqsa Hospital in Gaza in June and July 2024, describes what he found when he operated on patients:

“Very, very small entry wounds that sometimes you couldn’t even see the entry wound, but had significant damage and destruction internally. One evening a mother came in with a very small entry hole in her thigh — it had transected her femoral artery. At the exact same time, three of her children, ages one, four, and five, each had a single tiny entry tear in their abdomen, and each one had a bowel or colon injury that had to be repaired. There were no other signs of injury, no abrasions. Another girl, six years old, we could not find any wounds on her at all, and she had absolutely shredded blood vessels and bowel in her pelvis.”

Dr Firoze Sidhwa, a trauma surgeon who worked at Gaza European Hospital between March and April 2024, describes a similarly confounding pattern:

“One very, very small injury on the skin on his right chest and one on the left side — each one or two millimetres at maximum. No rib fractures. No other holes. A very strange and very unusual injury.”

What makes these injuries so devastating is the velocity at which the tungsten pellets travel. Dr Mark Perlmutter, an orthopaedic hand surgeon who worked in Gaza between March and April 2024, explains:

“These little tungsten pellets fly at such an impressively high velocity that they enter the skin and exit the other side of a limb with almost no visible wound, but the tissue on the inside is purely vaporised. Bone has so many fracture fragments that they’re impossible to fix. Muscle is turned to jello, arteries and nerves are thoroughly destroyed, and almost every injury that this causes requires an amputation.”

In two key ways, these projectiles behave like weapons that have already been banned under international law. They function like cluster munitions in that the explosive submunitions are unguided and not designed to engage a single target. And they cause superfluous injury and unnecessary suffering: injuries so severe and so difficult to treat that they present an extreme challenge even under optimal surgical conditions, let alone in a hospital that has been systematically starved of supplies, staff and electricity as part of the same genocidal war that deployed them.

The doctors’ testimonies also attest to the conditions under which this weapon has been deployed. Dr Mahmooda Sayed, a medicine physician who worked at Nasser Hospital between August and September 2024, noted that these injuries appeared consistently during times of evacuation — moments when people were necessarily gathered and in movement. Dr Tammy Abu Ghnaim, an emergency physician who also worked at Nasser Hospital between July and August 2024, describes what she witnessed:

“Most of the patients were paediatric, and they had been at a gathering of some kind. Very commonly, a group of people had gathered under the shade of a tree or canopy when they were hit, and then the cars that came to their rescue would also be hit. On Fridays, when the market was particularly busy, we would receive several groups of casualties from the market. We also saw incidents of people being targeted while waiting for taxis or buses. This was an almost daily occurrence, if not multiple times a day.”

At Nasser Hospital on August 25, 2025, rescue workers and journalists who had rushed to the scene of the first strike were struck by the second at approximately 10:17 am. This was a moment of congregation, exactly as Dr Abu Ghnaim describes. The acoustic signature is the same. The logic is the same. The pattern of the weapon’s deployment, against civilians in places of gathering, constitutes a war crime. The weapon itself, while not yet formally prohibited under international humanitarian law, may meet the criteria for prohibition.

What the evidence shows

The double-tap strike was once a term used to describe terrorist tactics: a method so condemned that states built entire legal frameworks to prosecute it. It is now state doctrine. The United States deployed it extensively in Pakistan, Yemen, and Afghanistan8. Israel has run it the length of its genocidal war on Gaza. What was once framed as the mark of a terrorist has become, quietly and without accountability, standard military practice.

From a single investigation: Three war crimes.

• The first is the double-tap strike itself. It is a tactic that works by design: the first strike draws in responders, the second kills them. In the case of Nasser Hospital, the sequence is visible on camera. Hussam al-Masri was hit by the first strike. The colleagues and rescue workers who ran towards him were hit by the second, nine minutes later. Under international humanitarian law, double-tap strikes constitute a war crime. Either the second strike is deliberately aimed at those responding to the first, or it is fired with complete disregard for the civilians it will predictably kill. There is no version of a double-tap that is consistent with the obligation to protect civilian life.

• The second is the striking of a protected medical facility. Hospitals are afforded explicit protection under international humanitarian law. Nasser Hospital was not an incidental target. It was struck twice, nine minutes apart, with a weapon designed to maximise lethal area effect, at a location known to be used daily by journalists and medical staff.

• The third is the deliberate targeting of journalists and press infrastructure. 5 named journalists were killed in a strike that Netanyahu initially called a tragic mishap9, before the Israeli army reversed course and asserted it had targeted and killed terrorists10. The two statements cannot both be true11. Since October 7, 2023, the Committee to Protect Journalists has recorded over 197 journalists killed by Israeli fire12. This is not a series of tragic mishaps. It is a systematic campaign to eliminate those who document what is being done.

Beyond these three war crimes, the practice that runs through them all may open the door to new legal ground. The double-tap strike on Nasser Hospital is the most widely documented use of the M329 to date. A weapon indiscriminate by design, deployed without discrimination: its effects alone satisfy multiple criteria for prohibition under international humanitarian law. The conditions of its deployment add further weight. Together, they constitute the evidentiary foundation from which a formal prohibition can be built.

These findings are not isolated. They are instances of a method that has run the length of Israel’s genocidal war on Gaza, extended into Lebanon, and is now visible in Iran13. What distinguishes its current use is the scale and the consistency.

This is the normalisation of a method designed to kill the people who respond to violence, and to silence those who survive to document it.

Optional paid support: If you’d like to support our work, monthly subscriptions are available at $5.00. Earshot is the foremost independent nonprofit organization producing sonic investigations with and on behalf of communities affected by corporate, state, and environmental violence. Your support enables us to conduct investigations, develop the field, and establish sonic evidence within accountability frameworks.

13

On the use of double-tap strikes in Gaza, see: +972 Magazine, July 2025, https://www.972mag.com/double-tap-israel-gaza-airstrikes-rescue/; on Lebanon, 2024, see: The World from PRX, November 2024, https://theworld.org/stories/2024/11/07/israeli-double-tap-strikes-in-gaza-and-lebanon-raise-serious-ethical-and-legal-concerns/; on Ramlet al-Baida, Beirut, March 12, 2026, see: Al Jazeera, https://www.aljazeera.com/news/2026/3/12/israeli-double-tap-strike-hits-displaced-on-beirut-seafront-kills-eight and The National, https://www.thenationalnews.com/news/mena/2026/03/12/israeli-double-tap-attack-kills-12-on-central-beirut-beachfront-where-civilians-sought-refuge/; on the Minab school attack, Iran, February 28, 2026, see: Middle East Eye, https://www.middleeasteye.net/news/exclusive-iranian-girls-killed-double-tap-strikes-minab-school; on Niloofar Square, Tehran, see: Reza Sayah and Murtaza Hussain, "Witnesses Describe Horror Scene After 'Double-Tap' Bombing Kills Over 20 at Popular Tehran Square," Drop Site News, March 3, 2026. Available at:

Drop Site News

Witnesses Describe Horror Scene After “Double-Tap” Bombing Kills Over 20 at Popular Tehran Square

Drop Site is a reader-funded, independent news outlet. Without your support, we can’t operate. Please consider becoming a paid subscriber or making a 501(c)(3) tax-deductible donation today…

Read more

6 months ago · 922 likes · 42 comments · Reza Sayah and Murtaza Hussain

Through forensic audio analysis of three recordings — a 6-minute video and two phone calls captured by paramedics during the assault — we reconstructed the attack minute by minute. What emerges is a sustained, two-hour assault: Israeli soldiers firing from fixed positions, advancing while shooting, and ultimately executing aid workers at close range.

This investigation, conducted with Forensic Architecture (FA) and with the Palestinian Red Crescent Society (PRCS), is published by Drop Site News and presented in the UK Parliament on 24 February 2025.

What this investigation reveals

This investigation took nearly a year — not because of external delays, but because of what the evidence demanded. We listened to these recordings hundreds of times. We counted every gunshot and analysed its constituent sounds — its supersonic bullet shockwave, muzzle blasts, and echoes. We timestamped, catalogued, and measured the intervals between each of these sounds. We traced echoes off concrete walls to locate shooters with meter-level precision.

No other organisation — human rights body or media outlet — has analysed these recordings with this level of forensic detail. Without this audio analysis, the full scope of what occurred would have remained obscure.

Our investigation establishes three critical facts:

Survivor testimony is an accurate and reliable documentation of what occurred. Every major element detailed by survivors through testimonies — soldiers’ positions, the coordinated advance while firing, close-range executions between the vehicles, the use of missile launchers — is corroborated through audio analysis.

This was a sustained military operation against clearly marked emergency vehicles and personnel. Our investigation documents an attack lasting at least 2 hours, followed by the burial of ambulances and destruction of evidence.

Soldiers’ names revealed. Audio evidence enabled us to name three soldiers who were involved in this crime. These findings should advance efforts toward accountability.

📄 Full investigation: For complete detailed analysis of methodologies, calculations, and witness statements used to reach our conclusions, read the full report co-authored with FA: Israeli Executions of Palestinian Aid Workers and Efforts to Conceal Evidence: Tel al-Sultan, Gaza Strip, 23 March 2025

Optional paid support: If you'd like to support our work, monthly subscriptions are available at $5.00. Earshot is the foremost independent nonprofit organisation producing sonic investigations with and on behalf of communities affected by corporate, state, and environmental violence. Your support enables us to conduct investigations, develop the field, and establish sonic evidence within accountability frameworks.

What happened on March 23, 2025

At approximately 4:00 am on March 23, 2025, a Palestinian Red Crescent (PRCS) ambulance driving along Gush Katif Road, in al-Hashahin, near Rafah in the south of Gaza, came under Israeli fire. Two drivers — Ezz El-Din Shaat and Mustafa Khafaja — were killed. A third PRCS volunteer, Munther Abed, was taken captive by the Israeli military and survived the attack.

At approximately 5:09 am, a second convoy arrived at the location of the first ambulance: 3 PRCS ambulances, one fire truck, and one Civil Defence vehicle. As the convoy came to a halt, Israeli soldiers opened fire.

Over the next two hours, Israeli soldiers fired at least 910 rounds — at least 789 directly at the aid workers.

Among the victims’ personal belongings, one mobile phone was retrieved. On it: a 6-minute video taken by PRCS first responder volunteer Refaat Radwan, capturing the moment the convoy came under Israeli gunfire at 5:09 am.

Two other PRCS aid workers — trapped in the assault — managed to call PRCS headquarters while the attack was ongoing. At approximately 5:13 am, Ashraf Abu Libda (PRCS first responder volunteer) called while under heavy gunfire. Nearly two hours later, at 6:55 am, Asaad Al-Nasasra (PRCS paramedic) called while surrounded by Israeli soldiers. Ashraf did not survive the attack. Asaad did. Both calls were recorded.

These three recordings — Refaat’s video and the two recorded phone calls — formed the evidential basis of our investigation.

We analysed the audio from these recordings and, in collaboration with FA, conducted detailed interviews with the two survivors. As the attack occurred during hours of little to no daylight, and as two of the three recordings are phone calls, audio analysis became a crucial means of reconstructing the attack and corroborating the survivors’ testimonies.

What follows is a minute-by-minute account of what audio evidence can establish when visual evidence is deliberately destroyed.

I. Multiple shooters firing simultaneously at aid workers

Munther Abed testified being held at gunpoint by Israeli soldiers, with “an M16 assault rifle pointed at my back”, witnessing soldiers “shooting directly at the convoy”. Our audio analysis confirmed his account.

We counted at least 910 gunshots across the three recordings. The majority — 844 shots — were captured in Refaat’s 5-minute-30-second video. The density of gunfire in this recording frequently exceeded 900 rounds per minute (RPM) — the highest rate of fire for a single M16 rifle in automatic mode.

At one point in Refaat’s video, 5 gunshots are audible within 67 milliseconds — the shortest possible time it takes an M16 to cycle and fire again. The irregular intervals between these shots confirm at least 5 shooters firing simultaneously. This represents only one moment in the attack. The remaining 839 gunshots in Refaat’s video, combined with survivor testimony, suggest many more shooters were present. Asaad testified there were “at least 12 soldiers.” Munther said “at least 30.”

Audio ballistic analysis of the 844 gunshots in Refaat’s video established that his camera was consistently in the line of fire throughout the period of audible gunshots. By identifying the supersonic shockwave made by the bullet flying past the recording device, followed by the muzzle blast in 789 out of 844 gunshots, we confirmed that 93% of all audible gunfire was fired towards Refaat’s camera and the emergency vehicles.

II. Timeline of the attack

1. 5:09 am — 5:13 am: Fixed firing position on an elevated sandbank

   Audio ballistic analysis and echolocation confirmed the soldiers’ initial positions on an elevated sandbank 38 to 48 metres southeast of the vehicles. Echolocation, in this context, is the process of using the propagation of sound waves and their acoustic reflections to forensically reconstruct a scene or event. The time delay, direction, and distortion of the echo reveal the size, distance, texture, and shape of objects in the environment. This process allows us to estimate the position of a sound source and its movements.

   The Israeli Army’s destruction and clearance of Tel Al-Sultan left very few structures standing. This fundamentally changed the area’s acoustic behaviour. In a dense urban environment, gunshot echoes overlap and become indiscernible from one another. Here, with most structures destroyed, the remaining surfaces produced clearly distinguishable echoes — paradoxically strengthening our ability to determine shooter positions with meter-level precision.

   In the first 4 minutes of gunfire, echoes are audible between 558 and 603 milliseconds after the muzzle blast. We identified a 2-metre-tall and 50-metre-wide concrete wall 130 metres south of Refaat as the only surface capable of causing these echoes. Echolocation confirmed the position of Israeli soldiers on an elevated sandbank southeast of Refaat.

   The interval between muzzle blast and echo remained stable throughout these 4 minutes, indicating the soldiers remained stationary on the elevated sandbank while firing continuously.

   This analysis confirmed the location where Munther testified to being held at gunpoint, surrounded by Israeli soldiers who opened fire towards the vehicles as the second convoy arrived.  

2. 5:13 am: Coordinated advance while firing

   In the final 1 minute and 30 seconds of Refaat Radwan’s video, echolocation and audio ballistic analysis indicated that the soldiers moved towards the aid workers at a walking pace of approximately 1 metre per second while continuously shooting.

   Simulation of the sound of a gunshot fired from a position increasingly close to Refaat. In the final 1 minute and 30 seconds of Refaat’s video the echo previously heard between 558 and 603 milliseconds after the muzzle blast increases gradually and is heard up to 779 milliseconds after the muzzle blast. The simulation illustrates how this increasing echo interval corresponds with the soldiers moving away from the 2-metre-tall concrete wall identified as causing these echoes, and towards Refaat’s position.

3. 5:14 am: Close-range executions

   At approximately 5:13 am, Ashraf called the PRCS headquarters. We confirmed that Ashraf’s phone call overlapped with the final 42 seconds of Refaat’s video by synchronising the sound of 109 gunshots across the two recordings.

   These gunshots coincide with audible voices of Israeli soldiers, heard across the left and right audio channels in Refaat’s video. This is the first time in the recordings where we hear voices speaking Hebrew at this level of intelligibility, and the first time we hear those voices distinctly separated across the stereo field, indicating that the Israeli soldiers were positioned to both the left and right of Refaat’s phone.

   In stereo recordings, proximity magnifies the positional differences captured by the left and right microphones. At a distance, sound waves hit both microphones at roughly the same time and level. However, once up close, any slight movement of the source causes larger differences in arrival time and volume between the two channels, making the movement through the stereo field sound more pronounced and distinguishable.

   Therefore, the presence of the soldiers’ voices — heard both louder and across both left and right recording channels — strongly suggests that Refaat had become surrounded by Israeli soldiers who were positioned in close proximity to him.

   The audible presence of Israeli soldiers’ voices coincides with Refaat and Ashraf’s final words announcing the arrival of the Israeli soldiers.

   Synchronisation of Refaat’s video and Ashraf’s phone call with audible voices of Israeli soldiers as well as Refaat and Ashraf’s final words.

   After Refaat’s video ends, Ashraf’s call continues for 52 seconds, capturing 24 additional gunshots. 8 of these contain new and distinctive muzzle blast echoes, characterised by 2 to 3 echoes in quick succession, arriving 20 to 140 milliseconds after the muzzle blast.

   We analysed these echoes and determined they were reflecting off surfaces just 3-24 meters from Ashraf’s phone. The only surfaces within this range were the emergency vehicles themselves — strongly suggesting shooters were firing from between the vehicles, within meters of aid workers.

   For 3 specific gunshots in Ashraf’s call, we used the positions of the ambulances relative to Ashraf’s location to establish even more precise shooter positions. The arrangement and orientation of the ambulances created a kind of architecture around the shooters, and by measuring the intervals in time between these echoes we could determine which ambulance the gunshots were reflecting off and in turn the most likely position of the shooter as they fired the shot. For one of these gunshots we could place the shooter 1 to 4 meters from Ashraf.

   These gunshots coincide with the last time we hear Ashraf’s voice, suggesting they are, most likely, the shots that killed him.

   Simulation of the sound of gunshots fired from in-between the emergency vehicles, revealing the position of Israeli soldiers as close as 1 to 4 metres to the aid workers.

   This evidence corroborates survivor Asaad Al-Nasasra’s eyewitness testimony that Israeli soldiers “came down [from the sandbank], got close to them [aid workers] and shot them from close range,” and “were walking between them [aid workers] and shooting”.

4. 7:13 am Sustained operation and burial

   Over two hours after the attack on the second convoy, Asaad Al-Nasasra made a phone call to the PRCS headquarters that captured at least 42 additional gunshots, vehicle movement, and an explosion.

   We identified the acoustic signature of this explosion as matching the 3-phased operation — ejection, ignition, propulsion — of a Spike LR guided missile. This weapon is fired from “vehicle-mounted” or “man-portable” launchers and is produced by Israeli defense contractor Rafael Advanced Defense Systems.1

   Asaad’s testimony described exactly this: “I saw a soldier on one knee, with the launcher on his shoulder. From the mound. (...) When the [UN] vehicle entered I was actually on the phone [with the Palestinian Red Crescent]. [They were] asking me if there were any cars moving, and I told them yes, the UN vehicle. As soon as I told him yes, they hit the vehicle”.

III. Names of Israeli soldiers

We enhanced the audio from Asaad Al-Nasasra’s phone call to the PRCS headquarters at around 7:13 am and identified multiple voices of Israeli soldiers speaking in Hebrew. Fragments of this speech were made intelligible, revealing the names of three soldiers involved in the attack: Elias (nicknamed Lalas), Yotam, and Amatzia.

A transcript of the recording reads as follows:

Clip 1:

00:05 – 00:07: Lalas did you finish?

00:07 – 00:10: Then go replace them, we will put the guns on them and you can bring their bags.

00:36 – 00:38: What’s happening Yotam?

00:38 – 00:42: [Unintelligible] you don’t drop cover. Go first over there, give cover, get their head stuck into something. [sic]

01:17 – 01:18: [sound of explosion, likely Spike LR missile]

03:36 – 03:38: [mechanical sound]

Clip 3:

00:32 – 00:37: Elias wait a second so that he can see you, the one passing with a [unintelligible].

00:44 – 00:46: I’m on [unintelligible] right now.

00:57 – 00:58: It’s Amatzia.

00:58 – 00:59: All of you come.

01:30 – 01:32: Let’s remove the [unintelligible].

01:42 – 01:48: [sound of engine revving and vehicle movement]

02:25 – 02:27: Did everyone already leave?

02:27 – 02:28: What?

02:28 – 02:29: Did everyone already leave?

02:29 – 02:31: Down here, I can’t [unintelligible].

Broader implications for humanitarian protection

International humanitarian law exists to protect precisely these workers: those who respond to crisis, who operate under emblems meant to guarantee safety. This case demonstrates how the entire framework of humanitarian protection is being undermined.

When soldiers fire 900+ rounds at marked ambulances and aid workers, advance while shooting, execute survivors at close range, and bury the evidence within hours of the attack, the framework of humanitarian protection has not simply failed — it has been deliberately violated.

What sound reveals

Through forensic analysis of over 900 gunshots, we built an auditory reconstruction of what happened that day — from the initial attack on the ambulances to the final close-range executions.

Sound, long overlooked in human rights research, proved to be indispensable. It revealed what was designed to be invisible: the position of shooters, their movements, the distance between weapons and victims, the names of soldiers involved in the attack. This case demonstrates not only what happened in Tel Al-Sultan, but what sonic investigation can contribute to accountability and human rights research.

Yet, beyond the forensic findings, it is the echoes of this incident that continue to haunt us — not just for what they revealed about this attack, but for what they revealed about Gaza itself.

The same destruction that enabled us to locate shooters with precision also marked the erasure of an entire acoustic environment. For two years, we have listened to incidents across Gaza. But here, we heard how the sound of Gaza has fundamentally changed. The acoustic environment itself bears witness to systematic destruction — not just of buildings, but of the possibility for certain sounds to exist at all.

This investigation documents one attack. The changed soundscape documents something more difficult to name: the erasure of an entire acoustic world.

Then something happened.

“And then from that complete silence, I heard this sound.” (Witness 05)

The crowd suddenly split down the middle of the boulevard, compelled by a sound they consistently described as an unidentifiable vehicle approaching at high speed. Serbian authorities deny any attack occurred. The Federal Security Service (FSB) claims the disorder began with fireworks. But 3,244 written statements and 15 in-depth earwitness interviews tell a different story.

“It was not just what we heard, it was more the feeling. That we felt it was something that made us run away from there. So, this is what I experienced (…) This feeling that something is happening, and we don’t know what is going on, it just made us run away. This feeling that I never felt before because I don’t know what is going on.” (Witness 08)

At Earshot, we specialise in investigating what cannot be seen — the acoustic dimensions of violence that visual evidence misses. This case presented a unique challenge: a sound that barely registered on recording devices, yet was experienced intensely by thousands of people. How do you document violence that leaves no trace?

Optional paid support: If you'd like to support our work, monthly subscriptions are available at $5.00. Earshot is the foremost independent nonprofit organisation producing sonic investigations with and on behalf of communities affected by corporate, state, and environmental violence. Your support enables us to conduct investigations, develop the field, and establish sonic evidence within accountability frameworks.

The methodology: listening as an investigative means

We partnered with the Centre for Research, Transparency and Accountability (CRTA) to gather first-hand witness accounts. CRTA collected 3,244 written statements from protesters present on the boulevard that evening. These statements described various acoustic characteristics — unfamiliar sounds, rapid approach, physical sensations — which informed our creation for a reference library of 26 sound samples spanning jet engines, fireworks, vortex cannons, and other potential sources.

Over two 3-hour sessions, we interviewed 15 witnesses who were positioned at different points along the boulevard. We asked them to describe not just what they heard, but what they felt. We played them the 26 sound samples, asking them to identify what matched their experience. What emerged was striking in its consistency. All 15 witnesses described:

• An unfamiliar sound they had never heard before.

• Something rapidly approaching from the northern end of Kralja Milana Boulevard.

• A physical sensation — not just auditory — that moved through their bodies.

• An overwhelming urge to move away from the center of the boulevard.

The silent vigil: what witnesses could hear

"Nothing was happening. We were just standing. It was quiet because of the 15 minutes that we were giving for the people that are not with us anymore." (Witness 03)

Backed by the FSB report, Serbian authorities claim the disorder began with fireworks. We sought to verify this claim. To do so, we first needed to establish whether witnesses could reliably distinguish between different sounds during the vigil — whether a firework or other explainable sound could have caused the mass panic visible in circulating videos.

We began by asking each witness to describe all the sounds they remembered hearing during the silent vigil:

“Nothing special. Sounds, the sounds in the dead silence were only from drones. And eventually from some phones, notifications on phones.” (Witness 13)

“Before the attack, it was all silent because I can hear someone cough maybe 200 metres away from me because it was so much silent. And then the attack itself.” (Witness 08)

These accounts reveal the crowd’s extraordinary attentiveness. Interruptions of the silence were registered acutely because of the collective effort to remain silent. The ability to discern and recall a cough from 200 meters demonstrates the exceptional acoustic condition these protesters collectively manifested.

From our analysis of 11 video recordings, we derived an ambient noise level of approximately 50dB during the vigil. Witnesses mentioned hearing coughs, phone notifications, drones, and popcorn machines — often noting how far away these sounds were. Most would have measured at or below the general ambiance: 35-45dB for a distant cough, 50dB for a quadcopter 50 meters overhead, 40-60dB for a nearby phone notification.

Though microphones can have broader sensitivity to sound pressure and frequencies than the human ear, the ear possesses far more sophisticated auditory processing. Our findings demonstrate that the crowd exhibited the capacity to recall and discriminate a wide range of sounds at varying intensities. This establishes the necessity of incorporating earwitness testimony when determining the presence of sonic weaponry — and the limits of relying solely on video recordings.

“Lead investigator: OK. And do you feel that there, as the sound got louder, it was easier to disassociate that sound from the one that the people could have been making.
Witness 07: Oh yeah, of course, of course, especially when it flew over us and that that was then I was I was sure something happened.”

During interviews, we systematically eliminated potential sound sources. Witnesses ruled out fireworks, quadcopters flying overhead, and stampede noises created by layering footsteps. In fact, witnesses clearly differentiated between stampede sounds and the attack. They also ruled out sirens, sine sweeps typical of sonic weapons, and even the vortex cannon sound we’d initially identified in our preliminary assessment.

Two witnesses identified motorcycles in the distance during the vigil — significant because it demonstrates their ability to distinguish between vehicular sounds and the vehicular quality of the attack they later experienced:

“That was very quiet. I heard just drones recording above us and 1 motorcycle that tried to go over, but it stopped and [was turned off] in the middle of the street.” (Witness 12)

“I saw two drones above our heads. And then, uh, you can hear the whirr of the drone. And then they were, I think gone.” (Witness 14)

1. An unfamiliar sound

   “The sound was very strange and alien to me. I don’t know if I can actually reproduce something like that.” (Witness 11)

   Having established witnesses’ capacity to identify sounds during the vigil, we turned to the sound of the attack itself. Eight witnesses explicitly stated they had never heard “a sound like that” before. All 15 witnesses, along with more than 60% of the written statements, used a combination of at least two sounds to describe what they heard — suggesting the attack was experienced as an unfamiliar auditory event.

   “Yeah. I think it wasn’t people because I felt something, like it’s something that you don’t see, but you feel it. And maybe I felt like, uh, it’s going to like, take us from the ground and put us on the floor. I felt something like that. It’s going to take me and I’m going to be on the floor, all of us.” (Witness 03)

2. An embodied sound

   “It wasn’t that much the sound. It was the feeling. The heart rates jumped up and me and people around me had this, you know, need to move away from the spot that we were in and this is it.” (Witness 09)

   For many, the experience wasn’t simply auditory — it was accompanied by intense physical sensation. In 2,335 written statements (72%), the sound came with such sensation. Eight witnesses described it as a force permeating or entering the body, prompting an instinctive reaction to move away from the center of the boulevard.

   “It was not just what we heard, it was more the feeling. That we felt it was something that made us run away from there. So, this is what I experienced (…) This feeling that something is happening, and we don’t know what is going on, it just made us run away. This feeling that I never felt before because I don’t know what is going on.” (Witness 08)

   “I felt like this sound was going through my body like I felt it in my bones, in my muscles. And throughout the head with the chills in my body as I felt the sound.” (Witness 09)

   “It’s like, it’s coming towards me (…) and then it was in me. You know it’s like that.” (Witness 11)

   This consistent description of sound as internal sensation provides a crucial clue. It’s unique to highly directional acoustic weapons to channel intense sounds in narrow streams, causing a disorienting sensation that witnesses describe as moving “inside” or “through” them. The experience was unfamiliar both in its perceived appearance and as an acoustic phenomenon — witnesses described hearing an unfamiliar sound that also behaved in acoustically anomalous ways.

3. A vehicular sound, approaching at high speed

   “I thought that 200 bikers, people on the motorcycles are coming towards us (…) My biggest Impression is basically that feeling that, uh, that fear that some machine is going to get over” (Witness 02)

   “And then I heard very weird sound like I was 100% sure that a vehicle is approaching (…) coming towards us.” (Witness 06)

   The sound approached witnesses in a highly directional way from the northern end of Kralja Milana Boulevard, from Terazje Square — emphasized by 11 of 15 witnesses, all confirming the same direction. Nine witnesses explicitly reported how the sound advanced rapidly, approaching them in a vehicle-like manner. 1,907 written statements (59%) described the sound as having a machinic, vehicle-like quality.

   “A sound I never heard before. It was a mixture of like a herd of horses coming towards me, but also with something mechanical, not as a car but something bigger and louder coming towards me. So, something in between.” (Witness 05)

   “I could hear the sound when, like, you know, when you start the engine, and you hit the gas pedal hard and that sound I could hear, that sound like approaching and becoming more distant. That is why I was 100% sure that some things ran through the crowd.” (Witness 06)

   “It was just a roaring sound of like a jet engine or something like that. It just went through the crowd.” (Witness 07)

   Seven of eight witnesses who heard a jet engine sample recognized it as similar in intensity and sonic profile. The comparison was rarely based on one sound alone — they responded to how the sample mimicked perceived movement: advancing from a distance, growing in strength, occupying vertical or overhead space. This corroborates numerous accounts of a fast-approaching directional force resembling aircraft or high-speed mechanical propulsion.

   "Lead investigator: So, we're going to play a couple of sounds and see if they were present.
   Witness 08: Yes, yes, yes, yes.
   Lead investigator: Like this, huh? Something like that. I see. And you felt it, like coming closer. Closer.
   Witness 08: Yes, coming closer in high speed."

4. The fear of what cannot be seen

   “You know how, but and it was so unreal, and you cannot explain it to your brain that you hear the car coming, but you cannot see.” (Witness 01)

   Of 15 witnesses, 11 described something rapidly approaching, prompting immediate reaction. Though all could identify where the sound originated, they couldn’t locate a visible source — explicitly stated by eight witnesses. This absence of visible source for such an intense physical experience created pervasive fear and disorientation.

   “Well, just the shock of it, it was, I don’t know, we’ve, as I said, we’ve been to a lot of protest and every time when it’s the end of the silence, you get this release because everyone starts making noise again and we, everyone knows what it means. Everyone knows why we are there. So, we were just waiting for that to happen and instead of that we got something that we didn’t know what it was.” (Witness 07)

   Contrary to the FSB report’s claim that the crowd showed no acute distress, five witnesses described immediate sense of imminent threat. Crucially, this fear didn’t stem from the sound’s volume but from its unfamiliar acoustic profile, marked directionality, and lack of identifiable source. Witnesses constantly described anticipatory dread — the sensation that something was approaching at high speed, suggesting impending projectile or vehicle impact.

   393 written statements (12%) explicitly mentioned fear, discomfort, unease, and impending danger. Though no physical impact occurred, the convergence of testimonies describing this anticipatory response suggests the sound was collectively interpreted as a threat.

   “So, the intensity is what stuck with me and the moment I felt it, it was like I felt it in my chest. Then the heart started pumping. And then and like, a panic (…) I felt anxiety in my chest” (Witness 04)

   “Mostly it was something that caused strong fear in me, like something I had to explain for.” (Witness 06)

   For some, the sound triggered something deeper — responses linked to past trauma. Two witnesses explicitly referenced memories of NATO’s 1999 aerial bombardment of Belgrade, noting acoustic similarities. Even without physical harm, the sound activated conditioned fear responses, functioning as a psychological weapon by reanimating latent or inherited memories of violence.

   “I thought of NATO bombing. (...) That fear and discomfort that I felt in my body, not the sounds. Because those rackets and everything you couldn’t hear it, you couldn’t see it. You couldn’t see when the bomb [hit], [it] felt like bombing.” (Witness 01)

   “I remember the similar sound when it was bombing of Serbia and I was in my hometown and one of these bombs went straight above my head (...) Or just because of the memory of the attack, it’s uncomfortable. The memory of the bombing of Serbia, which what made it uncomfortable.” (Witness 08)

   The consistency of these testimonies enabled us to categorise the sound according to four characteristics: highly directional, engine-like, embodied, and perceptually ambiguous. These characteristics would guide our reconstruction.

Reconstructing the sound

From these testimonies, we reconstructed the sound heard by witnesses. By layering elements witnesses consistently identified and applying spatial effects corresponding to the four acoustic characteristics, we created a plausible acoustic reconstruction.

This reconstruction doesn't replicate the original sound at its source. Rather, it produces a composite drawing from the timbral qualities and perceptual impressions of spatial movement that witnesses described. The significance lies in substantiating a mass auditory experience shared by thousands. The ability to reconstruct the sound testifies to striking consistency between our 15 interviewed witnesses and CRTA's 3,244 written statements.

Despite Serbian authorities’ denials that a sonic weapon was used, and despite our inability to identify verifiable traces in available recordings (addressed in Part 4 of our report), the cross-corroborative accounts from thousands of individuals form substantial evidence. This reconstruction approximates how witnesses experienced the sound — countering ongoing efforts to negate the protesters’ accounts of that night.

📄 Full investigation: For complete technical analysis, sound propagation modeling, and all 15 earwitness testimonies, read our full report: Sonic Attack on a Silent Vigil (EN/SR)

In an interview with N1 Belgrade about our investigation, the interviewing journalist mentioned he had been present on Kralja Milana that night. When he heard our reconstructed sound, he confirmed: this was what he experienced.

The evidence hidden in plain sight

Days after the vigil, Serbian parliament member Marinika Tepić revealed procurement documents showing the Ministry of Internal Affairs had requested seven LRAD 450XL devices. Ahead of a press conference where she made these documents public, she posted an image on X of what appears to be such a device mounted on a police vehicle near the National Assembly, captioned: “On the right - a gathering of students and citizens. Left - POLICE HOLDING A SOUND CANNON on a jeep near the National Assembly of the State of Serbia.”

The LRAD 450XL is a highly directional acoustic weapon that can transmit sounds at up to 145 dB and remain intelligible over 1.7 kilometers. Its unique emission pattern — a 30° cone with dramatic sound decay outside the beam — could explain why the attack barely registered on recordings while being intensely experienced by those in its path.

We simulated the LRAD’s propagation characteristics. A listener positioned directly in the beam 160 meters from the source would hear it as loud as a chainsaw (103 dB).

Move just 12° off axis, and it drops to the volume of a vacuum cleaner (75 dB).

At 18° off axis — outside the 30° beam — it becomes as quiet as normal conversation (60 dB), potentially indistinguishable from ambient noise on recording devices positioned at the edges of the boulevard.

The correlation between witness testimony and LRAD characteristics is striking. The four characteristics witnesses consistently described — unfamiliarity, marked directionality, embodied physical sensation, and absence of visible source — map precisely onto the LRAD 450XL’s technical specifications. The device’s narrow directional beam explains why witnesses identified the sound’s origin while unable to locate a visual source. Its capacity to generate intense pressure waves accounts for the physical sensations. And crucially, the narrow beam resolves the acoustic paradox: how thousands experienced an intensely physical sound that left minimal trace on recordings captured outside the beam’s path.

Why the reconstructed sound matters beyond Belgrade

The reconstruction of the sound of the attack serves as a sonic identikit of a fugitive sound, one that evaded capture on recording devices and is denied by Serbian authorities, despite an overwhelming body of witness testimony.

By presenting a reconstruction built by its earwitnesses, our intent was to give form to an experience lived by thousands. When 3,259 people describe the same anomalous experience in remarkably consistent terms — across 3,244 written statements and 15 in-depth interviews — that consistency itself becomes evidence. This case demonstrates how sound captures forms of violence that resist visual documentation — from slow violence and environmental degradation to acoustic weapons explicitly designed to undermine collective action. It challenges how we think about accountability for state and corporate violence.

Our investigation transforms protester testimonies into a document about power itself: how subjugation operates through division, and resistance succeeds through collective action. This is a story not only about a violent attack but about resistance, for two sonic forces confronted each other on the evening of March 15, 2025: one was an acoustic weapon — our investigation concludes it was highly likely the LRAD 450XL; the other was a punishing and enduring 15 minutes of collective silence manifested by hundreds of thousands in response to governmental corruption and neglect.

The protesters that night exercised their fundamental right to peaceful assembly, choosing collective silence as their form of political action. What authorities attacked was not simply individual bodies but this collective right itself — the right to gather and manifest political will through coordinated silence.

On that night, protesters were targeted by a weapon that turned the collective body into its singular component parts. Sound was weaponised by authorities to divide. This investigation — through the reconstruction of a sound experienced by hundreds of thousands — demonstrates how sonic documentation becomes an act of resistance through restoration: reassembling what division sought to fragment.

Optional paid support: If you'd like to support our work, monthly subscriptions are available at $5.00. Earshot is the foremost independent nonprofit organization producing sonic investigations with and on behalf of communities affected by corporate, state, and environmental violence. Your support enables us to conduct investigations, develop the field, and establish sonic evidence within accountability frameworks.

Subscribe now