Audio Ducking

Audio ducking is the technique of automatically lowering one sound, usually music or ambience, while another, usually speech, is playing, and bringing it back up when the speech stops. It's what lets a voiceover, recorded or generated with text-to-speech, sit clearly on top of a soundtrack, a podcast host talk over a theme tune, or a radio presenter speak over the end of a song. Ducking is one of the most common mixing tasks in video and audio production, a basic tool of podcasts, explainers and other speech-heavy formats, and one of the first that editing software automated.

Two Ways to Duck

There are two fundamentally different approaches, and they suit different jobs.

Sidechain compression is the live, signal-driven method. A compressor sits on the music track, but instead of reacting to the music's own level it listens to a second input, the "sidechain", fed from the voice. Whenever the voice crosses a threshold, the compressor turns the music down; when the voice stops, it releases. The attack and release times control how quickly the music dips and recovers. Sidechain ducking works in real time, which is why broadcast and live sound use it, but it reacts to every breath and cough, and its pumping can be audible.

Planned gain envelopes are the editorial method. Instead of reacting moment to moment, the software analyzes where speech occurs and writes volume keyframes onto the music: down before each phrase, up after it, with fades of a chosen length. Because the envelope is computed from the whole timeline, it can ignore short gaps, start the dip slightly before the first word, and stay down across a pause that is too short to be worth recovering in. The result is visible and editable as ordinary keyframes, which suits editorial work.

Auto-Ducking in Editors

Adobe brought automatic ducking to Audition in late 2017 and to Premiere Pro in April 2018, both through the Essential Sound panel. The editor tags a clip as music, chooses what it should duck against (dialogue, sound effects or ambience), sets the sensitivity, the amount of reduction and the fade length, and the panel generates volume keyframes on the music track. Because the result is ordinary keyframes, it can be adjusted by hand afterwards.

Blackmagic Design took a different route in DaVinci Resolve. Resolve 19 (April 2024) added a Ducker track effect to its Fairlight audio page, which lowers one track whenever another is active without a manual sidechain setup, and Resolve 20 (May 2025) added an AI Audio Assistant that builds a complete mix, evening out dialogue and setting music and effects levels against it. Digital audio workstations have long offered sidechain compressors for the same job.

Setting It Well

• Depth. The right amount of reduction depends on the music: a dense track with vocals needs more than a sparse ambient bed. Set it by ear against the quietest speaker, not the loudest, and check it on small speakers, where masking is worst.
• Timing. The dip should finish before the first word, not on it. Starting the fade a fraction of a second early keeps the first syllable clear.
• Recovery. Bringing the music back too quickly between sentences makes it pump; holding it down through short pauses sounds more natural.
• Frequency, not just level. Sometimes the music is loud enough but masks the voice in the same range. Cutting the music's midrange under speech, where much of a voice's intelligibility sits, can clear the voice with less overall reduction; this is a form of dynamic equalization.
• Overall loudness. Ducking changes the mix but not the delivery target. The finished program still has to meet its loudness and true-peak target.

Ducking in Automated Editing

For AI video editing tools, ducking is a natural fit for the planned approach. A system that already has word-level timing from speech recognition knows exactly when someone is speaking and can compute an envelope from the transcript rather than from a level detector, which avoids reacting to coughs, room noise or a guest's microphone bleed. It can also reason about the music itself: ending a track on a beat or phrase, or choosing where it swells back up, are editorial decisions that a sidechain can't make. In an agentic editing workflow, where the edit is stored as data as in video as code, the ducking depth becomes one more parameter in the edit plan, previewed and adjusted like any other, rather than an effect baked into the audio.

Ducking also matters for generative music. A model can produce a track of exactly the right length and mood, but it still has to be mixed under dialogue, and the ducking envelope is often what makes generated music sound placed rather than pasted on.