Microphones, Channels, Gain, Clipping, and Calibration

Normalization stops where it does for a reason of order. What it rescales is the recorded number, not the capture that produced it. By the time a scaling factor is applied, saturation has already flattened the peaks. The microphone has already blocked the frequencies it does not pass, and its response is already stamped into the spectrum. Reverberation and competing sound have already masked the source. Scaling arrives after all of that and moves the whole mixture together. Frame-dropping and clock-drift are damage to the alignment between channels, and no per-file scaling factor addresses alignment at all.

So the microphone is not a neutral opening to the pipeline. It is a directional, frequency-dependent transducer embedded in a room and an electronic chain. Placement, polar pattern, gain control, clipping, channel synchronization and device metadata can all become part of the learned decision, whether that was intended or not.

That is why absolute calibration, relative normalization and domain normalization belong side by side, before anyone reaches for a different architecture. They buy different things. A fix aimed at the wrong one leaves the capture exactly as it was.

A respiratory-audio classifier can look like a screening tool while hearing something other than the disease. The clearest published instance analysed audio from 67,842 individuals, 23,514 of them PCR-positive, recruited through NHS Test-and-Trace and the REACT survey. Not a pilot, and not a small one. Coppock and colleagues reported it in Nature Machine Intelligence in February 2024: audio-based classifiers showed no evidence of improving on a simple symptoms checker.

Unadjusted, the classifiers looked like a screening tool: ROC-AUC 0.846 [0.838, 0.854]. Then the authors matched on the confounders they had measured. “However, after matching on measured confounders, such as age, gender, and self reported symptoms, our classifiers performance is much weaker (ROC-AUC 0.619 [0.594, 0.644]),” they write. A large part of that apparent skill was being carried by things that were not the respiratory sound, in a cohort of 67,842 people. The only reason anyone knows is that somebody ran the matched comparison.

The uncomfortable part is that measuring the microphones would not have produced either number. Frequency-response deviation and self-noise can be reported in full, for every handset in the study, and every figure can be correct. Those figures describe transducers. They say nothing about what the classifier was tracking.

The step from ±2 dB to ±1 dB turned on a single word, and it was bought with hardware. Kardous and Shaw returned to the same four apps in 2016: “The initial study examined 192 apps on the iOS and Android platforms and found four iOS apps with mean differences of ±2 dB of a reference sound level measurement system. This study evaluated the same four apps using external microphones. The results showed measurements within ±1 dB of the reference.” The NIOSH Science Bulletin pins the conditions: “The NIOSH SLM app, when used with an external calibrated microphone, measured sound levels within ± 1 dB of the reference SLM over the testing range of 65 -95 dB SPL in our laboratory”, evaluated for compliance with type 2 requirements of IEC 61672/ANSI S1.4. The decibel came from a microphone bolted on, not from a better setting.

Calibrated, not sensitive, not well specified. Polar pattern, sensitivity, automatic gain control and calibration name four different parts of one capture chain. Two of them can sound alike and still demand different proof, in different units, from different people. So any write-up of that chain has to use these four words exactly.

Those three operations were left standing side by side earlier. What separates them is where the answer comes from.

Absolute calibration maps digital values to a physical reference such as sound pressure, and that reference lives outside the file. It is why the NIOSH condition was an external calibrated microphone rather than a better setting on the handset. It is why OSHA had to name the A scale and slow response in the regulation itself.

Relative normalization stays inside the recording. That does not make it vague; it is standardised to the decimal place. ITU-R BS.1770 defines the measurement in LKFS/LUFS. EBU Recommendation R 128, first published in February 2010 and revised in November 2023, recommends “that the Programme Loudness Level shall be normalised to a Target Level of −23.0 LUFS. Where attaining the Target Level is not achievable practically (for example, live programmes), a tolerance of ±1.0 LU is permitted.” Quality-control workflows are held to ±0.2 LU, with a true-peak ceiling of −1 dBTP. A target, a tolerance and a ceiling — and still a number referenced to full scale rather than to sound pressure.

Domain normalization stays inside as well, and is doing something else again. It reduces device variation for a model, at the risk of removing part of the task signal with it. Each of the three has to be shown separately. An operation that never leaves the recording, ±0.2 LU included, cannot tell anyone that the recorder was the thing being heard.

The workflow runs in one direction. Characterize the transducer — measure frequency response, self-noise, directivity, sensitivity and nonlinear range. Specify placement and geometry — record distance, orientation, mounting, channel order and synchronization. Set gain with headroom — test quiet speech and worst-case transients while monitoring automatic gain behavior. Preserve calibration metadata — store device, firmware, channel, unit and calibration history with each recording.

What you measured about the transducer does not stay in the measurement report. The assumptions baked into how you characterized it are consumed downstream, when you specify placement and geometry, and nothing later reopens them. Every step after the first one trusts the first one.

That leaves a single place where a mistake can still surface. The calibration metadata kept with the file, device and firmware version included, is the part of the characterization that reaches whoever asks the awkward question months later. Discard it after ingestion — the fourth failure mode on the opening list — and there is nothing left to check the first step against. Preserve it, and a repair remains possible. CHiME-6 could compensate for frame-drops and clock-skew across 32 microphones per session two years after the recordings were made. It could do that because the audio and its device structure had survived intact.

Robustness across devices was written into the task from the start. “The systems are expected to be robust to different devices”, says the DCASE 2020 Challenge Task 1 description. Then the challenge published the table that shows what robust costs.

Task 1A used TAU Urban Acoustic Scenes 2020 Mobile: 64 hours of audio, 23,040 ten-second segments, 10 acoustic scenes, 10 cities and 9 devices — real A, B and C plus simulated S1–S6. The full dataset spans 12 European cities, two of which appear only in the 11-device evaluation set. The device-wise results table for the baseline reads 72.8% on device A, 68.9% on C, 62.7% on S1, 61.7% on B, 58.2% on S3 and 54.6% on S2. Same model, same scenes. That is an 18.2-point spread across the devices seen in training. On device D, which the model had not seen, it was 22.8%. The baseline's overall evaluation accuracy was 51.4% (50.5–52.3), against 76.5% (75.8–77.3) for the top system, Suh_ETRI_task1a_3.

One mean accuracy reports 51.4% and hides every one of those numbers. Frequency-response deviation and self-noise, meanwhile, would have described nine well-made capture paths. They would have said nothing about the 18.2 points.

The format already has a published proposal behind it, and the exercise is easier if you use it. In 2018 Gebru and six co-authors proposed that every dataset travel with a datasheet, and argued it by analogy: “In the electronics industry, every component, no matter how simple or complex, is accompanied with a datasheet that describes its operating characteristics, test results, recommended uses, and other information. By analogy, we propose that every dataset be accompanied with a datasheet that documents its motivation, composition, collection process, recommended uses, and so on.” Motivation, composition, collection process, recommended uses: four headings, and a capture chain fits all four.

Write the steps out in order. List every transformation, including analog filtering, gain, the ADC, compression, denoising and resampling. Map confounders — which devices, rooms and placements correlate with labels. That is the question whose unmatched answer was 0.846 and whose matched answer was 0.619. Design counterbalanced collection, spreading target classes across devices and environments. Then define rejection rules.

The page is worth writing only if a stranger can use it. It earns its place when another team can pick it up and ask, without your help, whether the recorder rather than the sound is doing the work in your results. And the part that gets skipped is the last one. The page has to name the point at which a recording is too clipped, too unsynchronized, or too undocumented to keep.