The Accuracy Ceiling Is in the Microphone, Not the Model: Room Capture and Multi-Channel Concurrency for On-Premise Minutes
Why does the same ASR model hit 98% in one meeting room and only about 80% in another? The difference is usually not the model — it is the room capture. This article breaks down the real gap between three capture approaches, how pickup distance and room acoustics affect results, what dual omnidirectional mics plus bone conduction actually solve, and how to size concurrency from a single room to many.
The same ASR model hits 98% in one meeting room and only about 80% in another — and the difference is usually not the model. It is the capture.
Many evaluations put the entire budget on "is the model accurate", then discover at acceptance that even with the lexicon configured and hot words added, the error rate will not come down. The reason is simple:
A model can only work within the limits of the input audio. Poor audio cannot be repaired by any model.
Note: Figures marked as published values come from product specifications; performance levels are typical-environment reference values that vary with acoustic conditions, pickup distance and concurrency strategy — measure them on site.
The accuracy ceiling sits at the capture end
Start with a fact that is easy to miss: speech recognition is the one stage in the chain that cannot compensate for what came before it.
A model can fix "this word may have been misrecognised". It cannot fix "this sentence was never recorded clearly". The latter shows up acoustically as three problems:
| Audio problem | Acoustic cause | Fixable by a better model |
|---|---|---|
| Far-field attenuation | Speech energy decays with distance, SNR degrades | Essentially no |
| Reverberation residue | Room reflections contaminate the speech | Partially |
| Multi-channel crosstalk | Overlapping speech, channel mixing | No |
The first is the deadliest. As pickup distance increases, direct-speech energy drops sharply while ambient noise does not — so the signal-to-noise ratio falls. Raising sensitivity at that point amplifies noise along with speech: it sounds louder, and accuracy can actually get worse.
The conclusion: capture is the first variable in the whole chain. This also explains why "the same engine performs very differently in different rooms" — what differs is not the engine, it is the room.
The real gap between three capture approaches
Meeting-room capture comes in roughly three routes, and selection must distinguish them:
| Approach | Principle | Multi-speaker | On-premise fit |
|---|---|---|---|
| Dedicated hardware omnidirectional mic | Mic array + array signal processing | Strong, designed for people around a table | Native |
| Meeting-system mixdown | Reuses the existing system's mixed output | Medium, bounded by that system | Needs integration verification |
| Software local recording | Client software records local audio | Weak, single channel only | Generally unsuitable |
Dedicated hardware omnidirectional mics are designed for people seated around a table: sensitivity is roughly uniform in every direction, and combined with noise suppression and echo cancellation, multiple speakers keep stable pickup quality. The transcription side takes USB omnidirectional microphones through standard interfaces, and desktop hosts support plug-and-play with an external mic.
Meeting-system mixdown looks convenient — reuse the audio output of the existing meeting system. But mix quality depends entirely on that system: if it was built for human ears, its dynamic range and SNR may not be sufficient for a model. What must be verified is channel separation in multi-party scenarios, because a mixdown often folds several channels into one, leaving speaker separation with no usable input.
Software local recording is basically a non-starter for on-premise deployments: it captures a single channel, cannot separate speakers in a multi-person meeting, and depends on a client install.
Common failure modes
- Treating "can record" as "can transcribe": meeting-system audio is mixed for human ears, not for models.
- Testing only near the podium: the four corners and the farthest seat are the real operating conditions.
- Ignoring channel count: once channels are folded into one, speaker separation has no input.
Pickup distance and room acoustics
Pickup distance is the parameter most often asked about and most often answered vaguely. With a desktop host plus an external omnidirectional mic, effective pickup covers roughly 5 to 10 metres, which suits small and medium meeting rooms.
What if you need to go beyond that range? The correct answer is more pickup points, not more sensitivity. This is dictated by acoustics: speech energy decays with distance while noise does not, so gain amplifies both equally.
Room acoustics matter just as much:
- Reverberation: glass walls, smooth floors and large empty rooms reflect strongly, contaminating speech with its own echo.
- Steady-state noise: air conditioning, projectors, outdoor traffic — these have stable spectra and are at least tractable algorithmically.
- Transient noise: chair movement, door slams, phone vibrations — the most common cause of a whole segment falling apart.
Run two passes during acceptance: one with air conditioning and projector running (real conditions) and one in a quiet room (control). A quiet-room-only test is not a test — most meeting rooms sit around 40 to 50 dB of real background noise.
Dual omnidirectional mics + bone conduction: not more microphones
The common instinct on portable recorders is "add more microphones". But microphone count is not pickup quality — multiple microphones without array signal processing just capture several noisy copies of the same signal.
The effective approach is complementary pickup mechanisms:
| Mechanism | What it captures | Strength | Weakness |
|---|---|---|---|
| Air conduction (omni mic) | Sound waves | Wide bandwidth, naturalness | SNR collapses under heavy noise |
| Bone conduction | Vibration through solid surfaces | Intelligibility survives heavy noise | Narrow band, requires contact |
The bone-conduction value case is specific: when air conditioning, projector fans or outdoor noise have already destroyed the airborne SNR, the speech component captured by contact remains relatively intact. This is not "one more microphone" — it is an independent capture channel.
A personal recorder card, for example, uses a dual omnidirectional mic plus bone conduction design with a 5 to 10 metre pickup range, and syncs recordings automatically to a local server, forming a "one-tap record → AI transcript → knowledge base" chain — suited to mobile work, interviews and field research where fixed deployment is impossible.
From one room to many: sizing concurrency
Once capture is solved, the next question is compute and channels. There is a common trap here: treating "channels per box" as the only metric.
Real scenarios need two separate counts: multi-speaker separation within one room, and multi-room parallelism. The first is speaker separation inside one audio stream, the second is multiple audio streams transcribed in parallel — they consume compute very differently.
Three tiers by room scale:
| Tier | Use case | Typical configuration | Key metrics |
|---|---|---|---|
| Desktop host | Single small room, personal desk | 46 TOPS edge compute, 8GB memory, 128GB expandable storage | Latency under 1s (published) |
| Department server | R&D centres, finance branches, regional chains | 5 concurrent channels, 166 TOPS, 24GB high-bandwidth memory | End-to-end latency under 1.5s |
| Enterprise server | Government, defence, finance, medical and other high-compliance settings | 10 concurrent channels, 326 TOPS, 48GB high-bandwidth memory | RTF below 0.3, end-to-end latency under 1s |
Size by "rooms holding meetings simultaneously" and leave about 20% headroom for ad-hoc sessions. Running an enterprise server for a single room is waste; running desktop hosts for several parallel rooms means queuing.
Multiple rooms: audio aggregation and cluster control
Department-level servers support distributed audio aggregation over the network — audio from multiple rooms is aggregated onto one server for unified processing, with a web console for management. This avoids deploying a separate system per room and keeps transcripts and recordings centrally managed.
At larger scale, multiple hosts are managed through a management platform for cluster control: unified authentication, remote configuration and coordinated scheduling, with transcripts and recordings managed centrally. The published figure scales to 50+ channels of clustered concurrency.
There is an engineering issue here that is easy to overlook: timestamp alignment. If channels arriving from an aggregation network are not synchronised, speaker separation will fail on channel misalignment — this must be validated on site with an overlapping-speech recording, not judged from single-channel results.
Acceptance checklist
| # | Item | Method | Pass criterion |
|---|---|---|---|
| 1 | Pickup coverage | Record at four corners and the farthest seat | All pass |
| 2 | Noise tolerance | One pass with AC and projector on | Error rate does not degrade |
| 3 | Multi-speaker separation | Three-plus speaker overlapping recording | No whole-segment misattribution |
| 4 | Bone-conduction benefit | Compare channels under heavy noise | Intelligibility improves |
| 5 | Concurrency | Stress-test by actual room count | ~20% headroom |
| 6 | Audio aggregation | Multi-room networked test | Timestamps aligned |
| 7 | Cluster control | Multi-host unified console | Auth / config / scheduling work |
| 8 | Subtitles and playback | Inspect HDMI output | Speaker / time / original / translation |
Items 1 and 2 are the most commonly skipped: testing near the podium once, in a quiet room once, is the classic "passes acceptance, fails in production" combination.
Further reading: Beyond 98% accuracy: five engineering stages that decide minutes quality | Running meeting transcription on Ascend 310P | Choosing an on-premise ASR solution
FAQ
Q: Why does the same model deliver very different accuracy in different meeting rooms?
A: Because the accuracy ceiling is set by the input audio, and a model can only approach that ceiling, never exceed it. Far-field attenuation, reverberation residue and multi-channel crosstalk are audio problems that no amount of model capacity fixes. The greater the pickup distance, the stronger the room reverberation and the more people talking at once, the lower the usable signal-to-noise ratio and the higher the error rate.
Q: What is the difference between an omnidirectional and a regular microphone?
A: An omnidirectional microphone has roughly uniform sensitivity in every direction, which suits people seated around a table. A regular microphone is directional: it works well when close to one speaker but degrades sharply once people move or multiple people speak. Conference rooms also need noise suppression and echo cancellation — being able to record and being able to transcribe are different things.
Q: What is bone-conduction pickup for?
A: Bone-conduction pickup captures vibration through a solid surface rather than through air. Its value is intelligibility under heavy noise: when air conditioning, projector fans or outdoor noise have already destroyed the airborne signal-to-noise ratio, the speech component captured by contact remains relatively intact.
Q: What does a pickup range of 5 to 10 metres actually mean?
A: It defines the room size a single microphone can cover. With a desktop host plus an external omnidirectional mic, effective pickup covers roughly 5 to 10 metres, which suits small and medium meeting rooms. Beyond that range, speech energy decays while reverberation accumulates; raising sensitivity only amplifies the noise along with the speech. The correct response is more pickup points, not more gain.
Q: How do I choose the number of concurrent channels?
A: Size it by the number of rooms holding meetings at the same time. A single small room needs a desktop host; department-level multi-room parallelism calls for a 5-channel server; headquarters, government or medical environments running several rooms in parallel call for a 10-channel enterprise server. Multiple hosts can also be managed as a cluster with unified authentication, remote configuration and coordinated scheduling — the published figure is 50+ channels of clustered concurrency.
Q: How is audio from several meeting rooms aggregated?
A: Department-level servers support distributed audio aggregation over the network: audio from multiple rooms is aggregated onto one server for unified processing, with a web console for management. This avoids deploying a separate system per room and keeps transcripts and recordings centrally managed.
Q: Do the capture hardware and the transcription engine have to come from the same vendor?
A: Not necessarily, but interface compatibility and clock synchronisation must be verified. The transcription side typically takes USB omnidirectional microphones or local meeting-system audio through standard interfaces; if capture and transcription come from different vendors, verify audio format, sample rate and timestamp alignment, or speaker separation will fail on channel misalignment.
Q: How is data security handled for capture hardware in an on-premise deployment?
A: Recorder-class devices sync recordings to a local server rather than to the cloud, and transcription plus minutes generation run entirely locally so audio and text never leave the intranet. During selection, confirm that the sync target is a local server and that the sync channel itself is encrypted.