How Loud Should Music Be Under Dialogue?

Updated 2026-10-27·3 min read·Dialogue & audio

Why AI-video mixes go wrong

Most mixes on AI shorts fail one of two ways. Either the score and effects fight the voice, so lines get lost on a phone speaker, or the music is ducked so hard that every line has a hole punched around it. The second one usually comes from sidechain ducking: an automatic dip in the music whenever the voice is present. It sounds like a radio ad.

To get a real target instead of a guess, we took a five-minute dialogue-and-action stretch of a professionally mixed Chinese short drama, split it into a voice stem and a bed (music plus effects), and measured both.

What the reference actually did

• The whole mix sat around -14 LUFS integrated, with peaks pushed up near 0 dBFS. That's a loud, limited master, which suits phones and social feeds.
• There was almost no ducking. The bed measured within about half a dB whether or not anyone was speaking.
• During speech, dialogue sat about 6 dB over the bed. That one number is what kept the lines clear.
• The bed was allowed to be loud. Most of the runtime had no speech, and in those stretches the music and effects were simply the loudest thing playing.

The clarity came from where the bed was set. Nobody was moving it up and down. For comparison, a duck of around 5 dB under every line is far heavier than anything this reference used.

How to mix your short this way

1. Level the dialogue first. Every line from every character should land at a consistent loudness before you touch the music. Uneven voices are the usual reason people reach for heavy ducking.
2. Set the bed against a talking stretch. Play a section with dialogue and pull the music and effects down until the voice sits about 6 dB above them. A loudness meter on each stem makes this quick.
3. Leave the bed where it is. Don't ride it up between lines. If you must duck, keep it to a dB or two so nobody hears it pump.
4. Master the whole thing to about -14 LUFS integrated with a limiter. The reference ran its peaks right to 0 dBFS; leaving about 1 dB of headroom is a safer habit if a platform will re-encode your file.
5. Check on a phone speaker. That's where your audience is, and it's where a buried line shows up first.

A quick example

Say a 90-second episode opens on an argument over a tense string score, then cuts to a silent chase. Level both actors' lines first. Then, on the argument, set the strings so the voices clear them by about 6 dB. Don't automate the strings up for the chase. At the same fader position they already carry it, because nothing is competing with them there. Master to -14 LUFS and listen on a phone.

One caveat

This is one measured sample, five minutes long. Treat it as a good starting point that you can adjust. A pure-dialogue scene with no action may want the bed a little lower again.

In ShowMaker

Every character gets a fixed voice, so lines come in at a consistent level from the start. The editor holds the dialogue, sound and music score on the same timeline as the cut, so you set the bed against the voice right where you trim the scene. Start a series from the AI short drama generator, or read how we keep dialogue in sync with the picture.