Automatic gain control (AGC) in audio conferencing aims to produce an even speech output level, e.g., by amplifying weak speech and attenuating strong speech. A non-zero adaptation time is needed to reliably measure current speech level prior to applying gain, e.g., too fast an application of gain can result in over- and under-shooting, which manifests as unpleasant aural artifacts. When multiple voices of varying strength in a room take turns speaking, the AGC re-adapts frequently, resulting in audible and unnatural aural side-effects.

Techniques of this disclosure provide automatic gain control for each microphone, and apply the AGC gain after microphone selection but before mixing of microphone signals. Gains are adapted only across selected microphones, or microphones with energy levels similar to a selected microphone. In this manner, the techniques produce an even speech output level without the potentially jarring side-effects of AGC adaptation.