Blog
Two people are talking at your booth and the file remembers neither
Two conversations a metre apart arrive at one microphone as one signal, and the research corpus that documents this failure has a name for it that no trade-show article has ever used.
Two visitors reach the same counter within a minute of each other. One staffer takes the left seat, another takes the right, and for the next eight minutes two separate sales conversations run about a metre apart. That is a good hour on a show floor. It is also the exact condition under which simultaneous conversations at a trade show booth stop being recoverable afterward.
The recording succeeds. The record does not. What comes back is one file holding both conversations, layered, belonging to neither visitor.
What happens when two conversations reach one microphone
One microphone in a booth with two active conversations returns one signal containing both. Nothing in that signal marks which words came from which seat. A transcript generated from it is a single stream of text whose speaker labels drift across the counter, and the output cannot be attributed to either visitor with confidence.
Far-field diarization is the task of deciding who spoke when, from microphones placed away from the talkers rather than clipped to them. The distance is what makes it hard. A lapel microphone sits centimetres from one mouth, so the direct sound of that voice arrives far louder than anything else in the room. A microphone a metre or two out receives the direct sound, the reflections of that sound off the booth walls and the hall ceiling, the neighbouring conversation, and the ambient roar of the hall, all at comparable level. When two people talk at once, their speech overlaps in the same frequency bands at similar amplitude, and a system that assumes one active speaker at a time has no correct answer available. Research corpora such as CHiME-6, the sixth Computational Hearing in Multisource Environments challenge, and the AMI Meeting Corpus exist because this specific combination of distance, reverberation, competing sources and overlap defeats methods that work on clean single-speaker audio.
That is the failure a booth produces, and it produces it every time the booth is busy. A quiet booth records cleanly and teaches you nothing. The hour worth capturing is the hour with 2 conversations in it, and on a 3-day show that hour repeats.
The trade-show content world does not name this. Search the phrase and the results are arXiv preprints, conference proceedings and speech-toolkit documentation. Search for booth follow-up and the results are vendor blogs about workflow and note discipline, none of which mention overlap. Both bodies of writing describe the same afternoon. Neither cites the other. The gap between an empty note field and a lost conversation is partly a workflow problem, and partly this.
Why the research corpus calls this far-field diarization
The vocabulary belongs to speech research, not to event marketing. Diarization means segmenting audio by speaker identity, and far-field means the microphone is at a distance rather than on the body. The literature has treated this as an open problem for over a decade, with named corpora, published pipelines and reported error rates.
Mosner and colleagues gave the constraint its plainest framing at Interspeech 2018, in Dereverberation and Beamforming in Robust Far-Field Speaker Recognition. The title is the argument: at distance, you do not run speaker recognition on the raw signal. You run it on a signal that dereverberation and beamforming have already reshaped. Skipping those stages does not degrade the result gracefully. It changes what the downstream model is looking at.
That is the bridge worth crossing. A person shopping for booth capture is reading pages about CRM sync and badge scanning. The question they actually have, once two visitors sit down at once, is a far-field diarization question, and it has a fifteen-year paper trail nobody in the buying process has been shown.
What a research-grade pipeline has to do first
Published systems do not hand raw far-field audio to a diarizer. They spend most of their engineering upstream of it. The Johns Hopkins University submission to CHiME-6 is the clearest published account: a chain of enhancement stages runs before any speaker decision, and the diarizer’s output is then fed back into separation.
The paper describes the arrangement directly: “the beamformer outputs, one per array, are used for (i) speech activity detection and (ii) speaker diarization, both of which fuse information across arrays to improve accuracy, and (iii) overlap-aware variational Bayes hidden Markov model (VB-HMM) resegmentation to assign multiple speakers to overlapped speech regions. Speaker-marks from this diarization module are used in the multi-array GSS.”
| Stage | What it does | Why a booth needs it |
|---|---|---|
| Dereverberation | Suppresses the reflected copies of each voice | A booth is a hard-walled box beside other hard-walled boxes |
| Beamforming (BeamformIt) | Weighted delay-and-sum across array elements, steering sensitivity | Two seats are 2 directions, and direction is the only clean separator available |
| Speech activity detection | Marks where speech exists at all | Hall noise reads as speech to a naive detector |
| Diarization across arrays | Fuses evidence from multiple arrays into one who-spoke-when | One array’s blind angle is another’s clear one |
| Overlap-aware VB-HMM resegmentation | Assigns more than one speaker to overlapped regions | Simultaneous talk is the normal case, not the exception |
| Multi-array guided source separation | Uses the speaker marks to pull the streams apart | Separation is the last step, not the first |
Six stages, and the diarizer a buyer might imagine is the whole product appears at position four.
Read that column against a trade-show booth and the correspondence is exact. Every stage exists because of a condition the booth also has. This is the reason the distance a device can resolve matters more than any feature on a spec sheet.
Why the published work stops at four speakers
Multi-channel diarization work on the AMI Meeting Corpus, the Augmented Multi-party Interaction recordings used as a standard benchmark, scopes itself explicitly. The 2022 arXiv paper on multi-channel speaker diarization with spatial information targets meetings with no more than four speakers, uses the Steered-Response Power Phase Transform, SRP-PHAT, to estimate how many speakers are present, and fuses system outputs with DOVER-LAP.
That scope is a statement about where the method was evaluated, not a ceiling anyone has proven. It is also a warning. The published improvements are measured against other published systems on AMI’s evaluation sets, in meeting rooms, with a small speaker count and known recording conditions. None of that transfers to an exhibition hall by assertion. A hall has a noise floor a meeting room does not, an unknown and changing number of nearby talkers, and no control over what the neighbouring booth is doing with its own speakers.
The honest reading is narrow. The research demonstrates that spatial information from a microphone array helps, that estimating the speaker count separately helps, and that these methods have been characterised carefully within a bounded setting. It does not tell you what happens at a booth.
What fixed geometry changes, and what it does not
A microphone array bolted to a counter knows something a worn recorder never knows: where its own elements are, and where the seats are relative to them. Those angles do not change when a staffer turns, leans, or walks to the shelf. That is a genuinely better starting condition than a device swinging on a lanyard.
Known geometry is what makes direction usable at all. Beamforming works by combining channels with deliberate delays so that sound from one bearing adds constructively while sound from other bearings partially cancels. The delays are computed from the element positions. If the positions move, the beam moves with them, and a wearable’s positions move constantly. Fixing the array to the furniture removes that variable, which is one of several reasons the capture belongs in the counter rather than on a person.
Here is what fixed geometry does not do. It does not separate voices, and 2 known bearings are not 2 separated streams. It removes one source of error from a chain that still contains reverberation, hall noise, overlapping speech, and every downstream model’s own failure modes. Better conditions are not a result. A structurally favourable starting point is a reason to expect a better outcome, and expecting is not measuring.
The number we do not have
No measurement of N-way separation accuracy exists in our work, and this post will not state one. Not a percentage, not a diarization error rate, not a maximum number of simultaneous conversations. The measurement has not been made, so there is nothing to report and nothing to round.
This matters more than the paragraph it takes to say it. Everything above is either a citation to published research conducted by other people in other rooms, or arithmetic that a reader can re-derive. None of it is device performance. Fixed geometry gives known angles. Known angles are an input to a separation system, not an output of one, and no amount of favourable geometry substitutes for a number taken in a hall with real booths on either side.
The figures in the next section are of the second kind. They are requirements derived by calculation, which is why they can be published before anything has been measured: a requirement says what the hardware must not get wrong, while a performance figure says how well the finished thing works. Confusing the two is how a specification sheet becomes a claim. Anyone evaluating capture hardware should ask which kind of number they are being shown, which is the reason our vendor question list exists at all, and the reason we publish what this form factor cannot cover alongside what it can.
What we designed against instead
With no accuracy figure available, the design work is about not foreclosing options. Three constraints follow from arithmetic rather than from opinion, and each one describes a mistake that cannot be undone in software afterward.
Every microphone channel shares one clock. Sound travels roughly 343 metres per second, so 50 microseconds of timing skew between two channels is about 17 millimetres of apparent path difference. Across an array with elements spaced near nine centimetres, that resolves to roughly 10.7 degrees of bearing error, which is wider than the angle between two seats at one counter. Skew of that size arrives free with separately clocked USB microphones, where each device runs its own oscillator and drifts against the others over an 8-hour day. A common I2S or TDM bit clock and a synchronous ADC remove it at the hardware level, and nothing in the processing chain can recover it later. That is arithmetic setting a requirement, not a measurement of a built device.
Audio is stored at 48 kHz, multichannel, without lossy compression. One sample period at 48 kHz is 20.8 microseconds, about 7 millimetres of sound travel. At 16 kHz a sample period is 62.5 microseconds, about 21 millimetres, and the angular resolution available to a beamformer degrades with it permanently. Lossy codecs are worse than a lower sample rate, because Opus and AAC in multichannel mode do not preserve phase relationships between channels, and phase between channels is the entire signal a beamformer operates on. A file that sounds fine to a listener can be useless for separation, and the 3-fold storage saving is paid for once and permanently.
Processing order is fixed for the same reason. A diarizer such as pyannote sums multichannel input to mono internally, discarding the spatial information the array was built to collect. The order has to be preserve the original, separate spatially with a tool such as ODAS, downmix to mono at 16 kHz, and only then diarize and transcribe. Reverse 2 of those steps and a 4-channel array becomes an expensive single microphone.
These are constraints on a design, held before there is a result to defend, which is also how we approach the consent and notice question and why the indicator has no off switch. Two people talk at one counter for eight minutes. Whether both conversations survive that hour is not settled by a feature list. It is settled long before, by whether anything in the chain threw away the difference between left and right.