Device Bench

Device test · Audio in

Microphone Test

Start the test, speak, and watch the waveform and meter respond. Record a five-second clip and play it straight back to hear what the far end hears, or run the two-window measurement for a noise floor, a speech level, and the ratio between them. The readout separates the three faults people confuse — access blocked, no device at all, and a device that is open but silent. If more than one input is connected, pick it in the device list first: a headset mic, a laptop's built-in microphone and a USB microphone all appear in the same picker. Audio stays on this device; nothing is uploaded.

Only checking a microphone? You are on the right page. Checking both before a meeting runs the camera, the microphone and a playback check together and prints one summary.

Statusidle Inputnot requested Signal— RMS— Peak— Clipping— Sample rate— Channels— Noise floor— Speech— S/N—

Processing state is reported once the microphone is open.

Sample rate is reported once the microphone is open.

RMS and Peak are dBFS — relative to digital full scale, not loudness. A browser cannot measure sound pressure in dB SPL, so no number here is a volume in the room.

Two four-second windows, back to back. Stay quiet through the first so the page can measure the noise floor, then speak without pausing through the second; the difference between the two medians is the signal-to-noise figure. Nothing is recorded during either window — only the level is read, and only on this device.

The clip is captured into memory on this device and played back from there. Nothing is uploaded — there is no server to upload it to — and it is discarded when you press Stop or leave the page.

Playback comes out of your speakers or headphones. Turn the volume down before playing a clip back: if the microphone can hear the speakers the loop can build into a squeal, which is loud enough to hurt with headphones on.

How this test works

When you press start, the browser asks for microphone access and hands this page a live audio stream. The stream feeds a Web Audio AnalyserNode, and on every animation frame the page reads the most recent 2048 samples and derives three things from that one window: the RMS, which is the average signal energy and tracks how strong the input reads over time; the largest single sample magnitude, which is the true peak; and whether any sample in the window reached −0.1 dBFS, which is what the clipping counter counts. The waveform is drawn from the same window, so the trace, the meter and the numbers are three views of one measurement rather than three independent ones.

The record-and-playback loop uses the MediaRecorder API to capture five seconds into an in-memory buffer and play it straight back, which answers the question a meter cannot: what does the far end actually hear? The two-window measurement answers a different one. Four seconds of quiet establish the noise floor, four seconds of speech establish the speech level, and the difference between them is the signal-to-noise figure — a ratio measured on one scale in one session, which is why it survives the preamp gain that makes any single dBFS level meaningless on its own. Each window is reported as a median, the level exceeded by half the windows measured, so one door slam cannot manufacture a result; and a window that did not collect at least 80 readings at 20 per second is refused outright rather than reported at low confidence. Alongside all of it the Input field reports the capture state as the browser reports it — allowed, blocked, no input device, or in use by another app — and the Signal field separately reports whether audio is arriving at all, because a microphone that is allowed and delivering silence is a different fault from one that was refused. Ticking “ask for unprocessed audio” re-opens the stream with echo cancellation, noise suppression and automatic gain control all requested off, and the line under the controls then reports what the browser says it actually applied, which is not always what was asked for.

The limits are worth stating plainly. Every level here is dBFS — relative to digital full scale — so it describes how much of the available range the signal uses and says nothing about loudness in the room; a browser has no calibrated reference, so dB SPL is not something this or any other web page can measure. The sample rate shown is the rate of the audio graph reading the stream, and browsers frequently resample between the device and the page, so it is not necessarily the rate the hardware runs at; where the track does not report its own rate, the page says so rather than implying the two agree. Channel count appears only when the browser reports it. The clipping figure counts inspected windows rather than clipped samples, because consecutive analyzer reads overlap or skip depending on frame rate — the window is the unit this instrument can honestly stand behind. And the signal-to-noise figure is a description rather than a rating: automatic gain control lifts the quiet window between phrases and compresses the difference, so a figure measured with it on reads lower than the hardware deserves. Nothing on this page scores a microphone or passes judgement on a room, because a browser has no threshold to hold either of them to.

Frequently asked questions

Is my audio uploaded or stored anywhere?
No. The meter, the waveform and the recording all run inside your browser using the Web Audio API and MediaRecorder. There is no upload step and no server to upload to: the test clip is held in memory on this device, is never written to disk, and is discarded when you press Stop or close the page.
Nothing is happening — is my mic blocked, missing, or just silent?
Those are three different faults with three different fixes, and the Input and Signal fields tell them apart. Input reads blocked when the browser refused permission, no input device when the system reports no microphone at all, and in use by another app when something else is holding the device. If Input reads allowed and Signal reads silent, the microphone is open and delivering audio that never rises above −60 dBFS — which is usually a hardware mute switch, a muted headset button, or the wrong microphone selected in the picker.
Why does my recording sound flat, quiet, or unlike my voice?
The microphone recording test plays back the five-second clip the browser captured, and browsers apply echo cancellation, noise suppression and automatic gain control to microphone input by default, because most microphone use is calls. That processing flattens the very signal you are trying to measure. Tick “ask for unprocessed audio” to request all three off; the line beneath the controls then reports what the browser says it actually applied, which is not always what was asked for. Hearing your own voice also always sounds unfamiliar, since you normally hear it partly through bone conduction.
How do I fix "permission denied"?
Click the camera/microphone icon in the address bar and set microphone access to Allow, then reload. If you previously chose Block, the browser remembers it — you must change it in site settings; the page cannot re-ask on its own. Check the operating system too: on Windows, Settings → Privacy → Microphone; on macOS, System Settings → Privacy & Security → Microphone. A browser-level Allow cannot override a system-level block.
What do the level numbers mean?
RMS and Peak are both dBFS — decibels relative to digital full scale, where 0 is the largest value the format can hold. They describe how much of the digital range the signal uses, not how loud the room is: a browser has no calibrated reference, so it cannot measure sound pressure in dB SPL and nothing here is a loudness measurement. Normal speech usually reads between −30 and −12 dBFS on RMS. Peak holds the loudest single sample since you started or last reset, and Clipping counts the inspected windows that contained a sample at or above −0.1 dBFS. Noise floor and Speech are medians in dBFS from the two-window measurement, and S/N is the difference between them — a ratio in plain dB rather than a level. There is no pass mark for that ratio: it describes this session on this device, and automatic gain control compresses it by lifting the quiet window between phrases.

Related tests and guides