Decibel meter

How this decibel meter works, and why it says “estimated”

A sound level meter measures sound pressure: how far the air pressure swings above and below atmospheric as a sound passes. A browser never sees pressure. It sees numbers from your audio hardware, scaled so that the loudest value the converter can hold is 1.0 — levels in decibels relative to full scale, or dBFS. Which sound pressure a reading of −40 dBFS corresponds to depends on three things a web page cannot see: the sensitivity of your microphone capsule, the input volume set in your operating system, and whatever processing a driver or headset applies before the browser receives the signal.

So this sound level meter does the one honest thing available. It measures the dBFS level precisely and adds an offset to turn it into an estimated decibel figure. Until you calibrate, that offset is a round-number guess of 90 dB, and the meter says so on screen. Calibrate against a sound level meter or a decent phone app and the offset becomes the measured difference between the two — in principle what calibrating a professional meter involves too: present a known level, read it, correct the difference.

Before it listens, the meter asks the browser to switch off echo cancellation, noise suppression and automatic gain control. They are on by default because they make video calls sound better, and every one of them changes the level. Noise suppression removes precisely the steady background a noise meter exists to measure. Automatic gain turns a quiet room up and a loud one down until they read alike. The panel under the meter shows what the browser actually applied.

dB, dBA and the other letters on a sound level meter

Sound pressure level in decibels is twenty times the base-10 logarithm of the measured pressure divided by a reference of 20 micropascals, which is roughly the quietest 1 kHz tone a young, healthy ear can detect. Because the scale is logarithmic, the arithmetic is not what it looks like: +3 dB doubles the sound energy, +6 dB doubles the pressure, and +10 dB is conventionally heard as about twice as loud. Two identical fans running together read 3 dB higher than one, not double the number.

• dBA is A-weighted. The ear is less sensitive to low frequencies, especially at quiet levels, so the A curve defined in IEC 61672 turns them down before the level is summed: about −19 dB at 100 Hz, −9 dB at 250 Hz, no change at 1 kHz and a slight lift of about 1 dB between 2 and 4 kHz. The curve was first modelled on hearing at quiet levels, but it is now used for noise of every level, and nearly every noise guideline and workplace limit is written in dBA. This meter applies the curve to the frequency spectrum of each block of audio it analyses.
• Unweighted dB, sometimes written dBZ, counts every frequency equally. The gap between the unweighted and A-weighted readings tells you what kind of noise you have: a gap of several decibels means much of the energy is low-frequency hum, traffic rumble or ventilation.
• Fast is a 125-millisecond time weighting, the standard setting for noise that changes quickly; Slow, at one second, is its steadier sibling. The large number on this page is the Fast level.
• Leq, the equivalent continuous level, is the steady level that would carry the same total energy as the fluctuating noise over the whole measurement. It is what “average” means here, and loud moments dominate it: ten seconds at 80 dB and fifty at 50 dB average about 72 dB, not 55.
• Peak is the highest instantaneous value of the waveform, unweighted. It always sits above the Fast level, because Fast is an average of energy and peak is the tip of a single wave.

Typical noise levels, from a whisper to a siren

Published figures vary with the source and with how far away it was measured, so treat these as orientation rather than targets. They come from US public-health sources (the NIDCD and the CDC) and from surveys of restaurant noise.

• Whisper: about 30 dBA.
• Refrigerator hum, or a quiet room: about 40 dBA.
• Normal conversation: 60 to 70 dBA.
• Busy restaurant: roughly 70 to 80 dBA, and some are considerably louder.
• City traffic heard from inside a car: around 80 to 85 dBA.
• Motorcycles: 80 to 110 dBA.
• Sports events and concerts: 94 to 110 dBA.
• Sirens: 110 to 129 dBA.
• Fireworks: 140 to 160 dBA.

How loud is too loud? Hearing-safety figures

The US National Institute on Deafness and Other Communication Disorders puts the starting point simply: sounds at or below 70 dBA are unlikely to cause hearing loss even after long exposure, while long or repeated exposure at or above 85 dBA can.

The occupational figures build on that 85. NIOSH recommends a limit of 85 dBA averaged over an eight-hour working day, with a 3 dB exchange rate: every 3 dB louder halves the safe time, so 88 dBA is four hours and 100 dBA is fifteen minutes. The legally enforceable US limit, set by OSHA, is looser — 90 dBA over eight hours with a 5 dB exchange rate — and requires a hearing conservation programme from 85 dBA. In the UK, the Control of Noise at Work Regulations 2005 set action values of 80 and 85 dB(A) of daily or weekly average exposure, and an exposure limit of 87 dB(A) that takes account of any hearing protection worn.

All of these are exposure limits: level and duration together. Thirty seconds at 88 dBA is not a hazard; a working day of it is. This page is not a noise dosimeter, and uncalibrated it is not accurate enough for any compliance or health decision. Use it to notice that somewhere is loud, then measure it properly.

Why room noise decides dictation accuracy

Speech recognition hears everything the microphone hears. What matters is less the absolute noise level than the gap between your voice and the room, both measured at the microphone — the signal-to-noise ratio. The kind of noise matters as well. A steady fan or air conditioner is the easier case. The hardest is other people talking, a television or a podcast in the background, because it occupies the same frequencies as your voice and can end up in your transcript.

Distance is the lever you control. In open space, the level from a source falls by about 6 dB each time the distance doubles. Moving from a laptop microphone half a metre away to a headset microphone three centimetres from your mouth is four doublings, roughly 24 dB more of your voice at the microphone, while the room noise, which arrives from every direction, barely changes. That is why the verdict above says “use a headset” for a moderately noisy room rather than “give up”.

The bands in that verdict are our rule of thumb, not a published standard: below about 45 dBA any microphone copes; 45 to 55 is fine with the microphone reasonably close; 55 to 65 needs a headset; above 65, expect errors even with one. Measure for ten seconds with nobody talking, and remember the verdict is only as good as the calibration. To see the same noise in the audio pipeline’s own units, the microphone test reports your noise floor in dBFS alongside your speaking level. When the room checks out, try dictating into the browser with the voice typing tool.

How to calibrate the meter

1. Find a reference. A sound level meter is best. A phone app is the practical alternative: NIOSH publishes a free Sound Level Meter app for iOS devices, and reports that it is accurate to within ±2 dBA in its own laboratory testing. Set the reference to A-weighting and Fast if it offers the choice.
2. Put it next to your microphone. Right beside it, pointing the same way, so both hear the same sound field.
3. Make a steady sound louder than the room. A fan or running water works. Speech is a poor calibration source, because it never holds still long enough for two meters to agree.
4. Enter the reading. Once both have settled for a few seconds, type the reference figure into the calibration box, say whether it is dBA or unweighted, and press Calibrate. The meter compares your number with its own average over the last three seconds and moves every reading by the difference.

The offset is saved in this browser and belongs to one microphone at one input volume. Plug in a different headset or move the input slider in your system settings and the calibration no longer applies. A single-point calibration also assumes the microphone responds the same way at every level, which holds well in the middle of its range and less well near the extremes, so calibrate at a level close to the one you want to measure.

What a browser decibel meter cannot do

• Correct your microphone’s frequency response. Computer and headset microphones are built for speech and many roll off the low frequencies, so a deep rumble can read lower than it is. Calibration fixes the overall offset at the frequencies you calibrated with, not the shape of the response.
• Read a very quiet room. Every microphone has its own electronic noise. Below that point the meter is measuring the microphone, not the room.
• Read very loud sounds. When the signal reaches full scale it clips and the reading stops rising. The meter warns you when that happens; turn the input volume down and calibrate again.
• Undo processing it cannot see. Some headsets, Bluetooth links and operating-system “audio enhancements” process the signal before the browser receives it, and no web page can switch that off.
• Keep full pace in a background tab. Browsers slow down timers on hidden pages, so keep this tab in front while you measure.
• Replace a certified instrument. Sound level meters are classified under IEC 61672 and are periodically calibrated against a reference. For workplace assessments, planning conditions or a noise complaint, you need one of those.

It also checks the input side only. To test what comes out of your speakers or headphones, use the sound test.

Where VoiceSnap Pro fits

We built this meter because dictation succeeds or fails on the room before the software gets a say. VoiceSnap Pro is a voice-to-text app for macOS and Windows: hold one keyboard shortcut, speak, and clean punctuated text lands in whatever field your cursor is already in — an email, a Slack message, a code editor. Every dictation is also saved to a searchable notes library. It is a one-time $39 purchase, not a subscription.

It has not shipped yet, so there is nothing to download today. Join the waitlist for a single email on release day. In the meantime, check your input with the microphone test and see what dictated text reads like with the browser-based voice typing tool.