Set a target, cap how far the gain may travel, and the level rides toward that target as the file plays.
When it is the right tool
Three tools on this site change level and they are not interchangeable.
| Tool | Measures | Applies | Fixes |
|---|---|---|---|
| Normalizer | the whole file, in advance | one gain, everywhere | a file that is too quiet overall |
| Auto-Leveler | continuously, as it plays | a gain that moves over seconds | a file that is uneven within itself |
| Compressor | continuously, in milliseconds | a gain that moves per transient | peaks inside a phrase |
An interview where one person sat close to the microphone and the other sat back is the case this was built for. A normalizer cannot help, because there is no single number that is right for both voices. A compressor set hard enough to fix it will audibly pump on every word.
The controls
Target is an RMS level, not a peak level. Speech at -18 dB RMS typically peaks somewhere around -6 dBFS, so the headroom above the target is where the transients live. Setting a target of -6 leaves almost none.
Max lift and Max cut are the travel limits. They exist because an unconstrained AGC will find something to amplify in every recording, including the air conditioning. On clean studio material 6 dB either way is plenty. On phone recordings and lecture captures, 18 dB of lift is not unusual.
Response sets how long the gain takes to reach a new target. Below 200 ms the movement starts to be audible on sustained sounds. Above 2000 ms it stops tracking a speaker who changes distance mid-sentence. 800 ms sits where most speech works.
Gate is the level below which the gain stops climbing.
The detector reads the input
The gain is calculated from the level going in, not the level coming out. A detector that reads its own output forms a feedback loop, and a feedback loop with a slow time constant hunts: the gain overshoots, the detector sees the overshoot, the gain corrects past the target, and the level wanders in a cycle of a few seconds. Reading the input means the gain asked for at any moment depends only on the material, so the movement settles instead of oscillating.
Example: a 40-minute interview
Voice A averages -14 dB RMS, voice B averages -29 dB RMS. With Target at -18, Max lift 18, Max cut 12 and Response 800 ms, voice A is pulled down about 4 dB and voice B is lifted about 11 dB, both inside their limits. The gate at -50 keeps the gain from climbing during the pauses between questions, where the room tone sits around -58 dB RMS.
Limits
The gain applies equally to every channel, so the stereo image does not move. There is no lookahead and no latency, which is what makes it usable as a live preview step, and also why it cannot catch a transient that arrives before the detector has responded to it. That is the limiter’s job.
It also cannot separate two sources that overlap. If both voices are on one track and they talk over each other, the leveler sees the sum and levels the sum.