Mixing Dialogue, Music, and SFX Levels for Web Video

When music swells and buries a sentence someone on screen is trying to finish, you lose the thread, and your hand moves toward the back button before you even know why. The order is fixed. Dialogue is at the top, key sound effects are in the middle, music and ambience form the base, and every other sound has to work around the voice, not compete with it. A pyramid makes that priority order concrete: dialogue is paramount, sound effects and Foley build realism in the layer beneath it, and music and ambience form the emotional floor underneath both. This reflects how the brain sorts competing audio signals, and when the hierarchy breaks, comprehension breaks first: the rest of this piece is less a list of tips than an account of how that one structural idea plays out at every stage of a mix.

The four layers of a professional web video soundtrack

A professional web video mix runs at least four distinct layers, and each one has a separate job to do.

Dialogue and narration sit at the top of the stack. The audience has to hear and understand it without effort, and everything built into the layers beneath it exists to support that single requirement.

Key sound effects and Foley occupy the layer below. Transition effects live here too. Whooshes, risers, and impacts have genuine structural purpose, not decoration: a short whoosh under a cut conceals the edit and adds momentum, a rising-pitch sweep pulls the viewer's attention forward, and a low impact gives weight to a key line or statement at the moment it lands.

Ambience and room tone form the next layer. This layer establishes the environment and keeps the silence between lines of dialogue from reading as dead air. The fix is simple in practice: record 30 seconds of pure room tone at every shoot and lay it continuously under the dialogue track, and the cuts disappear.

Music forms the base of the stack. What a viewer hears shapes how they read what's on screen. If you pair the same footage with two different scores, you can get two entirely different scenes, and that's part of why music carries its own narrative arc across a piece, rising at the climax and easing off at the resolution in step with the editorial pacing.

The target numbers that enforce the hierarchy in practice

The hierarchy isn't only a conceptual ordering. Dialogue anchors the whole structure: a working target averages −12 dB to −10 dB, with peaks kept below roughly −6 dB, aimed at a final loudness near −14 LUFS for YouTube delivery. Dialogue also has a separate reference point of −18 dB to −9 dB, a wider bracket that overlaps with the first but isn't identical to it, so you can see these figures are starting ranges professionals calibrate against, not a single locked specification handed down from above. Pushing dialogue too quiet causes YouTube's normalization engine to boost the signal on playback, pulling up noise along with the voice. But if you push it too loud, the platform compresses the track flat, flattening the dynamic range the mix was built with.

Every other layer is set relative to that dialogue range, not independently of it. Sound effects run in a broader bracket, roughly −10 dB to −20 dB, calibrated to what the specific effect needs to do: a hard impact meant to land hits higher in that range, a subtle ambient detail sits lower, and anything pushed above −10 dB risks audible distortion. Room tone sits lowest of all, around −45 dB to −50 dB, so quiet that you never hear it as a distinct sound, and it only works as the glue that keeps the space between words and between cuts from feeling empty. At the master bus, you set the final loudness target for YouTube delivery at −14 LUFS, and you set a hard limiter ceiling at −1 dB true peak, so no stray transient clips when you export.

None of these numbers replace listening. They're the numerical skeleton the hierarchy hangs on, but even if a mix measures correctly on every meter, you still have to hear it on a real device before it's finished, a point the mixing sequence later in this piece returns to directly.

Frequency shaping and layer masking

Two tracks can measure at exactly the right dB levels on paper and still compete on playback, because loudness and frequency space are two different dimensions of a mix, and level meters only measure one of them. Frequency masking is the quieter version of the same hierarchy failure covered above: the numbers all check out, and the dialogue still gets swallowed.

Speech intelligibility, the consonants and vowel detail that make dialogue understandable rather than just audible, lives mostly between 300 Hz and 4 kHz. A subtle EQ cut in that range on the music track solves the problem directly: dialogue pops forward without the dialogue fader moving, because the music has been asked to step out of the voice's frequency space as well as its volume space. So here EQ works in service of the hierarchy, not just tone for its own sake.

The dialogue track benefits from the same logic in reverse. If you put a high-pass filter below roughly 80 to 100 Hz on the dialogue track, it strips out low-end rumble that adds nothing to speech clarity, and that's exactly the frequency room a music track's bassline wants to sit in. De-essing and de-reverb handle narrower problems inside the voice itself, and the correction has to stay invisible: over-applied noise reduction leaves a voice sounding hollow and processed, which trades one audible problem for another.

Order matters here as much as it does anywhere else in the chain. De-noise first, then EQ, then compression. The sequence isn't a stylistic preference: each step depends on the one before it having already been done.

Auto-ducking and keyframe automation: enforcing the hierarchy dynamically across the timeline

Setting the right levels and the right frequency space solves the hierarchy at a single moment in the timeline. But you have to hold that hierarchy across every moment in the edit, through every transition between a line of dialogue and a stretch of music, without riding the volume fader on every single clip by hand.

Side-chain ducking handles this structurally. Side-chain ducking automatically drops the music track's level whenever the dialogue track carries signal, turning "music fights voice" from a recurring manual fix into a rule the system enforces on its own. In Premiere Pro, Auto Ducking inside the Essential Sound panel applies exactly this once clips are categorized by role, dialogue or music, ducking is switched on, and Generate Keyframes is clicked: the panel then applies processing based on the declared function of each clip rather than requiring the editor to find and adjust every overlap by hand.

Manual keyframing gives up some of that automation in exchange for finer control over how a transition actually feels. For the reverse pattern, bring the music back down and the dialogue back up when speech resumes. This is the manual version of what Auto Ducking does on its own, and it matters because it gives the editor taste-level control over how fast a transition moves and how it's shaped, instead of whatever default curve the automation produces.

Smaller seams in the timeline need their own fix. A crossfade on a dialogue cut, even one as short as 100 milliseconds, keeps the join between two dialogue clips from registering as a hard splice. An Exponential Fade bridges that specific gap, and it restores a natural rhythm to the cut.

The mixing sequence that makes all the preceding decisions cohere

You only get a clean mix from every technique described above if you apply it in the right order. Done out of sequence, these same tools compound into problems that take far longer to untangle than they would have taken to prevent if applied correctly from the start.

Dialogue comes first, because every other decision in the mix is calibrated relative to the voice. Clean the source before setting its level: de-noise, de-ess, de-reverb, in that order, because setting a fader level on a noisy signal just locks the noise in at a louder volume once everything else is built on top of it. EQ and compression need to follow cleaning, never precede it, for the same reason you saw in the earlier section on frequency shaping.

Sound effects come second, so you place them relative to the dialogue level you've already set, not against an unset music track. You need to check transition effects, impacts, and ambient detail against the dialogue for masking, because they tend to sit in the same mid and upper-mid range where speech clarity lives.

Music comes last. With dialogue and sound effects already placed, the music's available headroom is whatever space remains, and that order structurally prevents the single most common hierarchy failure, music drowning dialogue, rather than requiring the editor to catch it after the fact. Where possible, match the music's BPM to the rhythm of the cuts: if cuts land every four seconds, the music's beat should land on those same cuts, because a mismatch feels visually off even to a viewer who can't articulate what's bothering them.

Mastering bus processing comes after all three layers are balanced against each other, not before: light compression to glue the elements together as one cohesive mix, followed by a hard limiter at −1 dB true peak as the final safety ceiling before export. Three listening tests close the process out, and if you skip them, that's where problems most often get left in a finished file. Play the mix back on a phone speaker: if dialogue is hard to follow or music feels overpowering there, the hierarchy has failed for the majority of viewers who will watch on mobile. Check the mix across more than one device. If a mix only sounds right through studio headphones, it isn't finished, so you can't skip the phone speaker test.

AI tools and editorial judgment in the hierarchy workflow

AI audio tools do their most valuable work on the technical groundwork of the hierarchy, noise removal, gain staging, automatic ducking, loudness normalization, freeing the editor's attention for the creative calls that decide whether a mix actually serves the story it's attached to.

iZotope RX functions as the industry standard for audio repair and restoration. Its Repair Assistant analyzes a track, identifies the problems in it, and proposes several processing chain options for the editor to audition and choose between, a guided process that puts broadcast-quality noise removal in reach of editors without a specialist audio engineering background, without requiring them to diagnose the underlying problem by ear first. That function slots directly into the de-noise step at the top of the dialogue processing chain you saw earlier.

Adobe Premiere Pro's built-in AI tools have narrowed the distance between a fully manual professional mix and an AI-assisted starting point. Enhanced Speech comes from Adobe Podcast technology, and it cleans dialogue directly on the timeline. None of this replaces the frequency shaping, keyframe refinement, or listening tests described in the sections above: what it does is shorten the time it takes to reach a clean technical baseline, leaving more of the session for the decisions that actually require judgment.

Where does the judgment still belong to the editor? Technical decisions, gain staging, frequency balance, basic level-setting, loudness targets, are now handled by AI at a standard that used to require a specialist engineer. Creative decisions, the vibe of a scene, tonal character, emotional pacing, where to place an automation ride by feel, remain the editor's call, with AI functioning as an assistant on all of them. The workflow that gets the most out of both: AI for the technical foundation, the editor in charge of everything that depends on taste. That same principle runs upstream of the mix. An AI platform that reads camera motion, pacing, and emotional tone in raw footage to produce a structured first cut hands the editor a timeline that already reflects intentional narrative choices rather than a random assembly of clips, and that intentionality is what makes the mixing decisions that follow more tractable: the story structure is already visible in the cut before a single fader gets touched.

How the hierarchy principle scales across different web video formats

The hierarchy itself doesn't change from one format to another. What shifts is its practical expression: the relative weight given to each layer, which loudness target applies, and how much work the sound effects layer is asked to do, all vary by format even while the underlying priority order, dialogue first, effects second, music and ambience beneath both, stays exactly where it was.

Short-form and social video compress the hierarchy without abandoning it. Fewer layers get placed, and each one is placed more deliberately, because in a fast-scroll environment the audio is doing a disproportionate share of the interpretive work: a subtle click makes an action read clearly, a short impact makes a reveal land with intention. Overloading a short format with effects pulls attention away from the message rather than reinforcing it, so restraint itself becomes part of the hierarchy discipline in this format.

Loudness targets shift by destination too, and that's a format variable rather than a fixed universal number: roughly −14 LUFS for YouTube, −16 LUFS for podcast distribution, −23 LUFS for broadcast under EBU R128. Knowing the destination before the mix begins is not a minor delivery detail you tuck in at the end. It sets the loudness ceiling the entire mix is built against from the first fader move onward: the hierarchy described throughout this piece is a framework an editor carries from project to project, recalibrating the same priority order against whatever format happens to be in front of them.