How the level is measured
Levels are in dBFS, decibels relative to full scale. 0 dBFS is the loudest sample your microphone's converter can produce, so every real level is negative: a quiet room sits somewhere around −70 to −50 dBFS, and clear speech a hand's width to an arm's length from the microphone reaches roughly −30 to −12.
The test reads two numbers from the same samples. RMS (root mean square) is the average energy, which is how loud something sounds; peak is the single largest sample, which is what decides whether it clips. Speech peaks sit 10 to 20 dB above its RMS, so a voice averaging −24 dBFS can still touch −6 on a hard consonant. The meter's bar shows RMS on a scale from -60 to 0 dBFS, green while comfortable, amber when loud and red close to full scale. The marker shows the highest recent peak: it holds for 1 second, then falls at 20 dB per second, like the peak light on a mixing desk.
Every sample is measured once. The audio arrives in your browser's audio engine at the microphone's sample rate, and on each screen refresh the test works out from the audio clock how many new samples came in and measures exactly those, so a slow frame can't count a sample twice. A stereo microphone gets one bar per channel.
What clipping is
Clipping happens when the signal is louder than the converter can represent. The tops of the waveform are cut flat at full scale, which sounds harsh and crackly, and nothing afterwards can restore what was cut off. The test counts any sample within 0.1% of full scale (±0.999) as clipped, lights the clip lamp for 2 seconds so a single one can't be missed, and counts each burst of clipping once. The waveform turns red where it hits the edges.
One clip on a cough or a laugh is nothing to worry about. If it keeps happening at your normal speaking level, the input volume (gain) is set too high. Turn it down in your system's sound settings, or with the gain knob on an audio interface, until your loudest speech stays out of the red.
Noise floor and signal detection
For the first 1 second after the microphone starts, the test measures the background while you stay quiet: the hum of a fan or a computer, room noise, and the microphone's own hiss. It takes a low percentile of the short-term levels in that time, so a click or a breath doesn't raise it. That is the noise floor.
After that, every 100 ms of audio is compared with it. Speech counts as detected when a window is at least 20 dB above the noise floor and louder than -55 dBFS, so a tap on the desk is too short to count and a faint sound in a silent room is too quiet. If every sample is exactly zero, nothing is arriving at all: the microphone is muted by a switch, by your system, or by a privacy setting, which is different from a quiet room.
Voice processing, and why calls sound different
Browsers normally run three kinds of processing on a microphone, the same ones video calls rely on. Echo cancellation removes the sound of your own speakers so the other side doesn't hear themselves. Noise suppression cuts steady background noise such as fans. Automatic gain control turns quiet voices up and loud ones down.
The test turns all three off by default, because they hide exactly what a microphone test should show: the real level, the real background noise and real clipping. Switch processing on to hear roughly what a call sends. Some browsers keep part of it on whatever the page asks; the test compares what it asked for with what the browser reports and tells you when they differ.
Reading the spectrum
The spectrum shows how loud the signal is at each frequency, from 20 Hz on the left to 20 kHz on the right, on a logarithmic axis so each octave gets the same width, the way hearing works. Its levels are per frequency band of an 8,192-point analysis, so they read lower than the dBFS meter. A thin spike at 50 or 60 Hz and its multiples is mains hum from a ground loop or a nearby power supply; a broad hiss across the top is electronic noise; a voice that stops dead around 4 or 8 kHz usually means a Bluetooth headset in call mode, which records at a low sample rate.
Your audio stays on this device
Nothing leaves your browser. The microphone is analyzed by your browser's own audio engine, and the 5 second test recording is kept in this tab's memory only, until you record again or leave the page. Stopping the test, switching microphones or navigating away turns the microphone off.
Questions
Why isn't my microphone detected?
Check that it's plugged in and switched on, a headset's plug is pushed fully in, and a Bluetooth headset is connected. Then check that your system sees it: on a Mac under System Settings, Sound, Input; on Windows under Settings, System, Sound, Input. On Windows, switching off microphone access in the privacy settings hides every microphone from browsers. If the test finds a microphone but the level stays flat at exactly nothing, it's muted, usually by a switch on the headset or in the system sound settings.
How do I allow microphone access?
Press Start and choose Allow when the browser asks. If you blocked it before, the browser stops asking. In Chrome or Edge, click the site controls icon at the left end of the address bar and set Microphone to Allow. In Firefox, click the crossed-out microphone in the address bar and remove the block. In Safari, open Settings for This Website. If the browser allows it but it still fails, your system is blocking the browser itself: on a Mac check System Settings, Privacy & Security, Microphone; on Windows check Settings, Privacy & security, Microphone.
What level should my microphone be at?
Speak as you would on a call. Your loudest words should reach around −12 to −6 dBFS peak, with the RMS bar mostly in the green and touching amber, and the background should sit below about −60 dBFS. If your voice barely moves the bar, raise the input volume or move closer. If the clip lamp lights at your normal level, turn the input down: a little quiet is easy to fix later, clipping is not.
Why does my voice sound different in calls?
Calls process your voice before sending it. Echo cancellation removes your speakers' sound, noise suppression removes steady background noise and can make a voice sound thin or watery, and automatic gain raises and lowers the level as you speak. The call is then compressed for the network. Turn voice processing on in this test to hear the first three; the recording plays back what the browser captured, before any network compression.
Why does my Bluetooth headset sound bad when I use its microphone?
Most Bluetooth headsets have two modes. For music they send high-quality stereo one way only; once an app opens the microphone they switch to a call mode that carries sound both ways at a low sample rate, often 16 kHz or even 8 kHz. Voices sound muffled, and the spectrum stops dead around 8 or 4 kHz. Newer headsets that support LE Audio or better call codecs do less badly. For recording, use your computer’s microphone or a wired headset and keep the Bluetooth headset for listening.
What is clipping, and how do I stop it?
Clipping is sound louder than the microphone's converter can represent: the tops of the waveform are cut off flat, which sounds harsh and can't be repaired afterwards. Lower the input volume in your system's sound settings, turn down the gain knob on an audio interface, or move a little further from the microphone. Automatic gain control also prevents it, at the cost of a level that moves around.
Is my audio uploaded or saved anywhere?
No. The microphone is analyzed by your browser on your device, and nothing is sent to a server. The test recording is held in this tab's memory until you record again or leave the page; it's only saved to your disk if you choose Save recording.