DeviceBenchFree online mic test
Mic test with a real meter, not a bouncing bar
Allow the microphone once and your voice moves a dBFS meter with a peak hold behind it, so you can see whether you land in the -18 to -6 window a call wants rather than guess at it. Five seconds of silence gives you a measured noise floor, three switches take the browser’s own echo cancellation, noise suppression and automatic gain off the signal, and a ten-second take plays back exactly what the far end receives. What no browser can reach is the hardware behind the driver: the capture API has no field for a capsule, a preamp or a serial number, so a page can describe your signal and never your microphone.
- 100% free
- No signup
- dBFS meter
- Noise floor in 5 s
- 10-second playback
Nothing is being read yet. Speech should peak between -18 and -6 dBFS once the input is open.
Nothing is open yet
Press Start. The request has to come from a click, because a microphone prompt that arrives before anyone asked for one gets dismissed on instinct — and browsers remember a dismissal.
Sit still and say nothing for five seconds. The page keeps about 150 readings and reports the middle one, so a single chair creak cannot set your floor.
This browser has no MediaRecorder, so the meter and the floor run work but there is nothing here to play back.
Everything above is read from the live track, which is the whole of a page’s view: a label, a rate, a channel count and three processing flags. The capture API has no field for a capsule, a preamp or a serial number, and no measurement anywhere in it separates a microphone that is quiet from one that is simply far away from your mouth.
How to test a microphone
Three passes: get the right input open, find your level, then listen to what it really sounds like.
Open the input, then make sure it is the right one
Press Start and answer the prompt. Then look at the picker: a laptop with a webcam and a headset attached usually offers four or five recording devices, and the browser opens whichever the operating system last called the default rather than the one near your mouth. Names are hidden from a page until you have granted access once, so if the list reads Input 1 and Input 2, allow it and the list fills itself in.
Speak at call volume and watch where the peak lands
Read a sentence aloud at the level you would use to talk to someone. The bar shows the level right now and the thin marker behind it stays at the loudest moment so far. Aim for peaks between -18 and -6 dBFS: below -18 the far end turns you up and brings your room with you, and above -6 there is no headroom left, so a laugh clips and the clip warning latches on.
Go quiet for five seconds, then record ten
Stop talking and press the floor button. The page keeps about 150 readings and reports the middle one, which is your room and your gain measured together rather than guessed at. Then take the ten-second clip and listen to it: a hum, a fan or a keyboard you long ago stopped noticing becomes obvious the moment you are listening instead of talking.
Technical specifications
| Meter | RMS in dBFS from a 1,024-sample time-domain buffer, published 30 times a second, drawn from -60 dBFS up to the full-scale ceiling |
|---|---|
| Peak hold | The loudest RMS since the last clear, kept as a marker on the bar; a separate clip flag latches when any single sample reaches full scale |
| Noise floor run | 5 seconds of your silence, roughly 150 readings, reported as the median and as the level the loudest 5% still exceeded |
| Floor bands | Under -60 dBFS quiet, -60 to -50 typical, -50 to -40 audible in every pause, above -40 something is feeding the input |
| Target for speech | Peaks between -18 and -6 dBFS; above -6 a laugh has nowhere left to go and clips |
| Recording | 10 seconds requested at 96 kbit/s, which lands near 120 KB — Opus in WebM where the browser offers it, AAC in MP4 on Safari — held in this tab's memory and never uploaded |
| Processing switches | echoCancellation, noiseSuppression and autoGainControl pushed onto the live track and read straight back, because a browser may grant a constraint it did not apply |
| Price | Free, no signup, and no cap on how many takes you record |
Frequently asked questions
The meter is flat while I am talking. Is the microphone dead?
Almost never — the usual cause is that a different input is open. A machine collects recording devices quietly: the laptop array, a webcam's built-in capsule, a headset boom, and on Windows a virtual Stereo Mix that records the output and therefore stays silent while you speak. Change the input in the picker above and watch the bar. If every device is flat, the next place to look is the recording level in the system sound panel, which some vendor utilities park at zero after an update.
My browser has stopped asking for permission at all.
There are two separate blocks and only one of them lives in the browser. The browser stores a refusal against this exact address, so reloading will never bring the prompt back: clear it from the crossed-out microphone at the left of the address bar in Chrome and Edge, from behind the padlock in Firefox, or under Settings for This Website in Safari. Above that sits the operating system's own switch — Privacy and security, then Microphone on Windows, System Settings, Privacy and Security, Microphone on macOS — and the macOS one only takes effect after the browser is quit and reopened, because that permission is read at launch.
Should I switch echo cancellation, noise suppression and gain control off?
For a call, no; for measuring the microphone, yes. Those three are on by default in every browser because they are what makes a laptop usable in a meeting: echo cancellation subtracts sound the browser itself played so the far end does not hear itself, noise suppression removes anything steady enough to look like a fan, and automatic gain rides your level up when you lean back. Each one also changes the signal you came here to judge — suppression flattens a real noise floor into a fake one, gain control makes a quiet microphone look fine until the room goes silent and the hiss comes up with it. Turn all three off, measure, then turn them back on before you join anything.
What counts as a good noise floor?
Under -60 dBFS is quiet, -60 to -50 is normal for a laptop array or a headset boom, -50 to -40 means the far end hears your room in every pause, and anything above -40 is something actively feeding the input. The number is only worth having with noise suppression off, and it is worth taking twice: once with the computer idle and once with it working, because a fan curve is the single most common thing hiding in the -45 band.
Why does the playback sound nothing like my own voice?
Because you have never heard your voice the way anyone else does. Speaking vibrates your skull, and that path carries low frequencies straight to the cochlea, so the version in your head has a bass weight that no microphone in the room can capture. What comes back from the clip is the honest one. Two smaller effects sit on top of the surprise: automatic gain will have lifted your quiet passages toward the loud ones, and you are hearing the result through your own headphones or speakers, whose character you stopped noticing years ago.
Is anything recorded while the meter is running?
No. The meter reads a buffer of samples, works out one number from it and lets the buffer go, thousands of times a minute, and nothing is accumulated anywhere. The ten-second take is the only thing on this page that keeps audio at all, it exists as a blob in this tab's memory, and it is replaced the moment you record again and gone when you close the tab.
Can this tell me whether the microphone or the cable is at fault?
Not on its own — a browser sees a driver, not the hardware behind it, so a broken conductor and a broken capsule arrive as the same silence. What settles it is a second input: open the picker, switch to the laptop's own array and speak again. If that one works, the fault is in the thing you plugged in rather than in the machine, and if the device in question is a headset the next page takes both cups and the boom apart one at a time.
About reading a microphone level in dBFS
Every number on this page is in dBFS, decibels relative to full scale, and it is a ceiling-referenced scale rather than a floor-referenced one: 0 dBFS is the loudest sample a converter can represent and everything real is therefore negative. That is why the useful target for a voice is a band and not a maximum. Peaks between -18 and -6 leave enough headroom for the moment you laugh or push a chair back, and cost you little at the other end, because lifting a signal recorded 18 dB down brings up 18 dB of a noise floor you have already measured, while a sample that hit the rail has thrown away the shape of the waveform and no amount of processing puts it back. Meters that fill a row of green blocks hide precisely this: they are drawn on a scale nobody defines, so “three quarters full” means one thing on one site and something else on the next.
The floor measurement is worth more than the level for diagnosing a bad-sounding microphone, and it is the reading almost no browser tool takes. Sit in silence and whatever the meter still shows is the sum of the room, the cable, the preamp and the converter, and it sets the ceiling on how good you can ever sound: a voice peaking at -12 over a -35 floor has 23 dB of separation, which is audibly noisy, while the same voice over a -60 floor has 48 dB and sounds clean. This is also where noise suppression earns its keep and ruins the measurement at the same time, which is why the switches here are switches rather than a fixed default. If the floor rises and falls with what is on your screen rather than with what is on your desk, the source is coil whine from a display or a power supply feeding into an unbalanced cable — park a monitor on the black screen and then on the white screen and watch the number move.
Where a browser stops is worth stating plainly. It hands you a driver’s view: a label, a sample rate, a channel count and three processing flags, with no way to reach past that to the capsule. It cannot tell a microphone that is quiet from one that is simply far away, and it cannot see a hardware mute on a cable or a case, because that cut happens before the sound card and leaves a perfectly healthy track carrying silence. If the device you are testing is a headset, the headset test walks the two cups and the boom as one unit and catches the faults that only a combined device has; if you are about to join a meeting rather than diagnose anything, the video call test runs the input alongside the camera, the speakers and the line. And if you are sizing a recording rather than judging one, the bitrate calculator turns kilobits per second into megabytes per hour.
What happens to the sound this page hears
Every number on this page is worked out by JavaScript running in the tab you are reading it in. Nothing you type, paste or open is uploaded, logged or kept, which is also why the tools carry on working after you disconnect from the network.
The meter and the floor run consume buffers of samples and discard them frame by frame — there is no accumulating array anywhere, which is why the floor result is a median of readings rather than a recording of your silence. The ten-second take is the one piece of audio that exists at all, and it stays raw recorded data the player reads straight from this tab — overwritten by your next take, gone on reload.