Sidechain Ducking
Sidechain ducking is automatic level control in which one audio signal turns another one down. A compressor is placed on the signal that should get out of the way — usually a music bed — but its level detector listens to a different signal, the sidechain or key input, usually the voice. Whenever the key rises above a set threshold, the compressor attenuates the music by an amount governed by the ratio; when the key falls back below the threshold, the music returns to full level. Attack and release set how quickly each of those happens. The result is a bed whose level tracks the speech continuously instead of being set once: it dips under every word, including words nobody anticipated, and swells back in the gaps. It is called ducking because the music ducks out of the way of the voice, and it is the standard way to put music under narration in film, broadcast and online video.
How It Actually Works
A compressor has two jobs that are usually performed on the same signal: a detector measures the level, and a gain stage turns the signal down when that level is high. Sidechaining separates them. The gain stage still operates on the music, but the detector is fed a copy of the voice. Nothing about the compressor changes — only what it is listening to.
Threshold. The level the key must exceed before any gain reduction happens. It has to sit above the noise floor and below ordinary speech. On a decent microphone in a quiet room the floor sits around -50 to -60 dBFS and conversational speech averages -25 to -15 dBFS, so a threshold in the region of -30 dBFS separates them cleanly. Set it too low and breaths, keyboard noise and room tone hold the bed down permanently; too high and only shouted words duck anything.
Ratio. How much of the key's excess above the threshold is converted into attenuation. A hard-knee compressor reduces gain by (key level − threshold) × (1 − 1/ratio) dB. With a threshold at -30 dBFS and a voice averaging -18 dBFS, that excess is 12 dB: at 4:1 the bed drops 9 dB, at 12:1 it drops 11 dB, at 20:1 it drops 11.4 dB. The curve flattens fast, which is why ratios above about 10:1 are effectively "as much as this key can produce".
Attack. How long the compressor takes to apply the reduction once the key crosses the threshold. A few milliseconds — normal for a drum compressor — chops the front off every musical note the instant a word begins, and that discontinuity is audible as a click or a suck. Several hundred milliseconds leaves the first word of each sentence sitting on top of an undipped bed. For speech, 50-300 ms is the working range and 150-200 ms is where most ducks land.
Release. How long the compressor takes to let go after the key drops back below the threshold, and the parameter that decides whether a duck is invisible or maddening. Gaps inside a spoken sentence — between words, around plosives, at commas — typically run 100-300 ms. Gaps betweensentences run from about 500 ms upward. A release around 500-600 ms therefore keeps the bed down through a sentence and returns it only at a genuine break. Set it to 100 ms and the music surges several times per sentence; that is the pumping people describe as the music "breathing".
Knee, detection and link. A soft knee begins reducing gradually as the key approaches the threshold instead of switching at it, which hides the onset. RMS detection averages the key over a short window so a single plosive does not slam the bed, where peak detection reacts to the transient. On stereo, linking on the average of both channels rather than the loudest avoids the image drifting when one side is hotter. These three are the difference between a duck you notice and one you do not.
Hold, where it exists. Dedicated broadcast duckers add a hold time that keeps full attenuation for a fixed period after the key falls silent, before the release even starts. Compressors without a hold — including ffmpeg's sidechaincompress — ask the release to do both jobs, which is the real reason ducking releases are set so much longer than mixing releases.
Why a Static Music Level Never Works
The obvious fix for music covering a voice is to pick a quieter number for the music. It fails for three independent reasons, and they compound.
Speech is not one level. Within a single unprocessed take a voice routinely moves 15-20 dB between a thrown-away aside and an emphasized line. A bed set to sit correctly under the loud delivery is on top of the quiet one. There is no number that is right for both, because the thing it is meant to be relative to keeps moving.
The gaps are the point. Music under narration exists to fill pauses and carry momentum across cuts. A static level holds the bed at its ducked level through every pause too, so it is inaudible exactly where it was supposed to work. Turning it up to fix that puts you back on top of the voice.
The edit keeps moving. Cut a sentence, drop in B-roll, speed up a passage, or lengthen a music span, and every hand-placed volume point is now in the wrong place. A level-driven duck has nothing to move — it responds to whatever audio ends up under it.
| Static gain | Gated step duck | Sidechain duck | |
|---|---|---|---|
| What sets the level | One number, chosen once | A list of speech windows | The key signal's level, continuously |
| Reacts to unplanned sound | ✗ | Only if it is in the transcript | ✓ |
| Level during pauses | Identical to under speech | Full, stepping back up | Rises smoothly, timed by release |
| Needs a transcript | ✗ | ✓ | ✗ |
| Typical failure | Buries the voice, or is inaudible | Audible steps; misses breaths and laughs | Pumping when attack or release is wrong |
| Cost to build | Nothing | One automation point per window | One compressor per music item |
The middle column is worth understanding rather than dismissing, because it is what most automatic ducking in transcript-based editors really is: a fixed attenuation switched on across the spans where the transcript says someone is talking. It is precise about words and blind to everything else — a laugh, a cough, an unscripted overlap — and its edges are steps rather than curves. It is also the only one of the three that can duck the music before a word starts, because it knows in advance where the word is.
Why It Matters in Practice
A 22-minute interview, two people, one music bed under the whole thing. You set the music by ear against the intro, where the host is projecting, and it sits nicely 14 dB down. Eight minutes in, the guest says something quietly and leans back from the mic, and the bed is now level with them. You lower the music 6 dB. Now the intro has no music in it, the pauses have nothing in them, and the two-second beat you left after the guest's best line — which the music was supposed to hold — is silence.
Fixing that by hand means volume automation on every passage, and then re-doing it after every cut. Ducking fixes it structurally: the bed plays at its full level, and the voice — whichever voice, at whatever level, wherever the edit later moves it — takes 11 dB out of it while it is speaking and gives it back half a second after it stops.
This is also why "the music is too loud" and "my video is too quiet" need separating before either is acted on. The first is balance and ducking is the answer. The second is level, and the answer is loudness normalization — which lifts the voice and the music together and changes their relationship not at all.
How to Set a Duck That Works
- 1Set the depth you want, then find the ratio that produces itDecide the target first — 10-12 dB down is the usual answer for a bed under narration. Then read the arithmetic backwards: depth = (key level above threshold) x (1 - 1/ratio). If your voice averages 12 dB above the threshold, no ratio on the dial will produce a 15 dB duck, and you need a lower threshold rather than a higher ratio.
- 2Put the threshold between the noise floor and the speechPlay a passage with the speaker silent and note the level of the room. Play a normal spoken sentence and note that. The threshold goes between them, nearer the room tone than the speech — around -30 dBFS on a clean recording. Breaths triggering the duck means it is too low.
- 3Slow the attack until the duck stops being audibleStart near 150 ms. If the first word of a sentence rides on top of the music, shorten it; if each word puts a dent in the music you can hear as a click or a suck, lengthen it. The goal is that the bed is already down by the second syllable and nobody can say when it moved.
- 4Set the release longer than the gaps between wordsStart near 500 ms, which is longer than the 100-300 ms gaps inside a sentence and shorter than most sentence breaks. Listen for the music surging mid-sentence — that is the release being too short. Listen for the bed never returning during a real pause — that is it being too long.
- 5Check it on a quiet passage, not a loud oneA duck set against confident delivery almost always under-triggers on the quietest speech in the programme, which is where the voice most needs the room. Find the quietest thing anyone says and make sure the bed still moves for it.
One setting normally serves a whole programme. If different passages genuinely need different depths, the problem is usually the recording's level consistency, not the compressor.
Common Mistakes
The threshold is on the voice, not on the music
This is the single most common misreading. Turning the music down does not stop it ducking, and turning it up does not make it duck harder: the detector never sees the music at all. Depth in dB is set entirely by how far the key signal rises above the threshold. If the duck is too deep, lower the ratio or raise the threshold; the music fader is a different control answering a different question.
Ratio is not depth
Ratio is a proportion, not an amount. Ignoring the knee, gain reduction is (key level above threshold) x (1 - 1/ratio), so a voice sitting 12 dB above the threshold ducks the bed by 6 dB at 2:1, 10.8 dB at 10:1 and 11.4 dB at 20:1. Past about 10:1 the extra ratio buys almost nothing, which is why raising the ratio to fix a shallow duck usually fails — the key is simply not far enough above the threshold.
Sidechain ducking is not the EDM pumping effect, but it is the same device
The four-on-the-floor pump is a kick drum keying a pad with a fast attack and a release tuned to the tempo, so the effect is meant to be heard. Ducking for speech uses a slow attack and a release longer than the gaps between words precisely so the effect is not heard. Same filter, opposite intent — which is why copying a producer's sidechain preset onto a podcast bed sounds wrong.
Ducking is not a fix for the wrong music
A track with a busy midrange, a vocal hook or a constantly moving arrangement competes with speech in the frequency range speech lives in. Ducking lowers its level; it does not move it out of the way spectrally. If the bed still fights the voice at 12 dB down, the track is the problem, not the compressor.
In ffmpeg, threshold is a linear amplitude, not a dB value
The sidechaincompress and acompressor filters take threshold as a number between roughly 0.00098 and 1, where 1 is full scale. So threshold=0.03 means about -30.5 dBFS, and threshold=0.125 — the default — means about -18 dBFS. Reading 0.03 as "-3 dB" and typing 0.5 to make it gentler sets the threshold to -6 dBFS, at which point nothing ducks at all.
Ducking is a balance tool, and loudness is a different problem
"The music is too loud" and "the video is too quiet" arrive in the same sentence and have nothing to do with each other. Raising the master lifts the music and the voice together and changes their relationship not at all; that is a loudness question, answered by normalization. Ducking answers the first complaint only.
A key that is not clean ducks on the wrong things
If the sidechain is the whole program mix rather than an isolated voice, then laughter, a chair scrape, a door and any music already baked into the recording all push the key above the threshold. The compressor is behaving correctly; it is being told the wrong thing is speech.
How This Works in Valmera
Ducking is on by default and has no dialog. Ask for music and the bed is added at -18 dB with smooth ducking already enabled, which renders as one ffmpeg sidechaincompress per music item, keyed to a copy of the program audio, at threshold 0.03 (about -30.5 dBFS), ratio 12, attack 180 ms and release 550 ms. Against a voice averaging -18 dBFS that produces roughly 11 dB of reduction — conventional bed depth, reached over about a fifth of a second and released about half a second after the talking stops.
The older behaviour is still available as a mode: a flat -12 dB applied only across the speech spans read from the transcript, with adjacent spans merged so the gating does not chatter. It is the gated step duck from the table above, and it is occasionally the better answer — on a video whose non-speech audio is loud enough to hold a level-driven duck open. And when no speech survives under the music window at all, music is added as the lead audio at -4 dB with no ducking, because there is nothing for it to get out of the way of.
What it does not do. There are no threshold, ratio, attack or release controls and no gain-reduction meter — ducking on or off, the two modes, and a gain, and that is the whole surface. The key is the program audio, not an isolated voice stem, so laughter, a door and a chair scrape duck the bed exactly as speech does, and music already baked into your recording holds the key above the threshold permanently; Valmera cannot separate music from speech in a single mixed track. A voiceover ducks the program audio underneath it, but with a flat -12 dB step across its own window rather than a sidechain. Sound effects never duck at all, by design — a whoosh that dips under the word it is punctuating is not an accent.
Related Terms
- LUFS and loudness normalization — level, as opposed to balance; the other half of every audio complaint.
- Silence detection — the same threshold-above-the-noise-floor problem, solved for cutting instead of mixing.
- Word-level timestamps — what a transcript-gated duck uses to decide where speech is.
- J-cut and L-cut — the other craft technique built on audio and picture not changing at the same instant.
- Transcription and ASR — where the speech spans come from, and how wrong they can be.
- Edit decision list — where a duck setting lives as a decision rather than as baked audio.
- The full glossary — every term, with the numbers that apply.
Frequently Asked Questions
Valmera is an agentic video editor: you describe the outcome and an AI agent performs the edit, including the mix and the export, which is rendered from your original file at source quality.
Put Music Under Your Voice Without Drawing a Single Volume Point
50 free credits on signup, no card. Upload your footage, ask for a music bed, and listen to the pauses.
Start free →