Automatic Gain Control

Automatic gain control is the century-old idea that a system should ride its own volume knob. Done crudely, it's the reason camcorder clips breathe with a rising whoosh of room noise between sentences. Done well, it's invisible — your level simply stays right. The difference is whether the loop knows what a voice is.

What AGC is

An automatic gain control is a feedback loop wrapped around an amplifier. It measures the signal's level, compares it against a target, nudges the gain toward that target, and repeats — the electronic equivalent of an engineer's hand resting on a fader. That framing matters: AGC's job is not to shape the sound but to keep the whole system operating at a sensible level, no matter how the source drifts. A talker leans back from the mic, swaps a headset, gets animated — the loop quietly restores the working level. Where gain staging is the discipline of setting levels correctly once, AGC is the mechanism that keeps them correct over time.

Born in radio, raised in telephony

AGC predates almost everything else in the audio processing toolbox. In the 1920s, AM radio listeners had two chronic annoyances: distant stations faded in and out as the ionosphere shifted, and adjacent stations on the dial arrived at wildly different strengths, forcing a grab for the volume knob with every retune. The fix was automatic volume control (AVC), developed in 1925 by Harold Wheeler at the Hazeltine Corporation: a circuit that derived a control voltage from the strength of the received carrier and used it to wind back the amplification of the receiver's earlier stages. Strong signal, less gain; weak signal, more. By the mid-1930s essentially every home AM set shipped with AVC, and the listener's hand finally left the knob.

Telephony adopted the same philosophy for a different pain: long-distance lines had a narrow usable window between noise at the bottom and overload at the top, and automatic level regulation helped keep speech inside it across circuits of unpredictable loss. Between them, radio and telephony established AGC's identity — a slow, level-keeping servo at the input of a system — decades before anyone put one in front of a computer microphone.

AGC is not compression

Because both adjust gain automatically, AGC and dynamic range compression are constantly confused, but they differ on two axes. The first is timescale. A compressor reacts in milliseconds, gripping individual syllables and consonants; its gain changes are fast enough to alter the envelope of the waveform, which is why compression has a sound. AGC works over seconds. It isn't trying to shape syllables — it's trying to find the right long-term operating point, and its adjustments should be slow enough that nobody can point to the moment they happened.

The second axis is what gets controlled. A compressor processes the waveform after capture. A true AGC controls the input — often literally the analog gain of the hardware preamp ahead of the converter. That distinction has a hard consequence: if the preamp gain is set so high that the converter clips, no downstream processing can repair it, because the information is gone before any software sees it. An AGC that turns the actual input down prevents clipping at the source in a way no plugin can. The two are complements, not rivals: AGC parks the level in the right neighborhood, and the compressor does the fine sculpting within it.

The classic failure: noise pumping

AGC's bad reputation comes from one specific, avoidable mistake. A naive loop has no idea what it's listening to — it just chases its target level. So when you stop talking, the loudest thing at the microphone becomes the room itself: the HVAC, the fan, the street. The loop dutifully raises gain, and the noise floor swells toward the target until you speak again, at which point the gain slams back down. On cheap camcorders and laptop capture chains this produced the signature noise pumping: silences that whoosh upward, speech that begins with a lurch. A generation of users learned to hunt through settings and disable "automatic microphone adjustment" entirely.

The failure isn't in the concept; it's in the detector. The loop was answering the wrong question — "is the signal at target?" — when the right question is "is the speech at target?" Silence contains no speech, so silence should contain no gain changes. Notice, too, that the failure compounds: the extra gain applied during a pause means the first word after it arrives amplified, sometimes clipped, and the loop then spends the opening of your sentence recovering from its own mistake. A loop that had simply held still would have been right the whole time.

Speech awareness: the modern fix

Modern AGC designs make the loop deaf to everything that isn't a voice. The core technique is noise floor tracking: continuously estimate how loud the room is when nothing is happening, typically by watching the quietest levels seen over a recent window. With a floor estimate in hand, the loop can demand that a sound stand well clear of it — and clear an absolute minimum level — before treating it as speech worth measuring. Everything else is ignored: the loop freezes rather than chases. The threshold has to be relative rather than absolute, because rooms differ — a floor-relative test that works in a quiet study still works next to an air conditioner, where a fixed threshold would mistake the machine for a mumble. The remaining refinement is asymmetry. Gain reductions happen fast, because the thing they prevent — clipping — is irreversible; gain increases happen slowly, because creeping upward is safe and sudden boosts are audible. The result is a servo that acts only on evidence of a voice, in the direction of caution.

In DeskBroadcast

DeskBroadcast's Auto-level is a speech-aware AGC in exactly this tradition, and it rides the hardware input gain — in deliberate 0.7-second ticks — rather than processing the waveform, so the signal reaching the converter is right to begin with. It tracks the room's noise floor as the quietest recent tick over a window of about 8.4 seconds, and it counts a moment as speech only when its peak is at least 12 dB above that floor and louder than −42 dBFS. When speech peaks push past −6 dBFS it pulls gain down quickly; when speech settles below −20 dBFS it creeps gain up slowly; and during silence it does nothing at all — which is precisely why it never amplifies background noise. The target is speech peaks around −12 dBFS, a healthy level with real headroom, starting from the baseline set by mic calibration and leaving the millisecond-scale work to the compressor further down the chain.