Every level the noise test shows is in dBFS: decibels relative to full scale. It is the unit of digital audio, and once you know where its zero sits, the numbers read easily.
Full scale is the ceiling
A digital audio sample is a number with a fixed range. In 16-bit audio, the kind RFC 7874 uses for its examples, it runs up to ±32767. Full scale is the largest value the format allows, and it is defined as 0 dBFS. Nothing louder exists inside the signal, so every real level is a negative number: −20 dBFS is quieter than 0, and −60 dBFS is quieter still.
When a mic sends more than the format can hold, the samples cannot go any higher, and the tops of the waveform are cut flat. That is clipping, and on a call it sounds like crackle or a voice breaking up. The noise test flags it when your peak goes above −1 dBFS or when more than 10 samples hit the ceiling.
How the numbers work
dBFS uses the decibel formula for signal level: 20 × log₁₀(level ÷ full scale). That makes it logarithmic. Halving the level takes away about 6 dB, and dividing it by ten takes away 20 dB:
| Level | Share of full scale | Where you meet it |
|---|---|---|
| 0 dBFS | All of it | The ceiling; nothing can go higher |
| −1 dBFS | About 89% | Where the noise test flags clipping |
| −6 dBFS | About half | The top of the peak range RØDE suggests |
| −12 dBFS | About a quarter | The bottom of that range |
| −22 dBFS | About 8% | The speech level WebRTC asks software to aim for |
| −60 dBFS | 0.1% | The test’s line for a good noise floor |
| −90 dBFS | About 0.003% | The self-noise of a mic built to Microsoft’s recommendation |
The WebRTC figure comes from RFC 7874, which sets the target as an RMS value of 2600 on the 16-bit scale; 20 × log₁₀(2600 ÷ 32767) is about −22. The −90 figure is Microsoft’s recommended sensitivity for built-in mic arrays, −26 dBFS for a 94 dB SPL tone, minus its recommended signal-to-noise ratio of 64 dB.
So −20 dBFS and −60 dBFS are not close. They are 40 dB apart, a factor of 100 in level. Speech in the low −20s over a floor at −60 dBFS is a comfortable gap, and signal-to-noise ratio puts a number on it.
Peak and RMS
Two readings matter. Peak is the single largest sample, the one that clips first. RMS (root mean square) is the average power over a short window, closer to how strong a sound stays. The test’s live meter shows both. A pure tone’s RMS sits about 3 dB below its peak, because a sine’s RMS is 0.707 of its peak. The test’s Peak reading tells you about clipping; its Noise floor and Speech readings are RMS.
dBFS is level, not loudness
dBFS tells you how big the signal is. How loud a programme sounds has standards of its own. ITU-R BS.1770 sets out how to measure programme loudness and true peak, and the EBU’s R 128 recommends an average programme loudness of −23 LUFS, loudness units relative to full scale. Even WebRTC’s speech target is not a plain dBFS reading: RFC 7874 says to measure it using only frequencies above 300 Hz, whatever the sample rate.
Why apps even out your level
If everyone on a call sent whatever level their mic produced, you would spend the call turning people up and down. RFC 7874 says as much: a standard level for speech saves listeners from adjusting playback and makes it easier to mix voices in a conference. Its advice is that software that doesn’t control the whole capture path should use automatic gain control to hold active speech at the target, within 6 dB either way, and should not adjust music or desktop sharing automatically.
That is what the switches in your apps do. Zoom’s Automatically adjust microphone volume keeps a consistent level. Teams’ Automatically adjust mic sensitivity is there so everyone hears you at a level volume. Slack huddles and Steam voice chat have automatic gain control too. In the browser it is the auto gain setting, and the noise test reports whether your browser applied it.
The trade-off is that gain control reacts to what it hears. TeamSpeak describes it as lowering a strong signal and raising a weak one, so in your pauses it can lift the room. Zoom suggests turning its setting off if your audio fades in and out, and Valve suggests turning Steam’s off if friends say you randomly get louder and quieter.
Reading your own numbers
In the noise test, aim for speech within about 6 dB of −22 dBFS, the WebRTC target; the test calls anything under −30 dBFS too quiet. Keep peaks under about −6 dBFS, and the floor at −60 dBFS or below. If your speech reads too low, move closer before you turn anything up. If your peaks hit the ceiling, turn the gain down.