Music under speech

Music under speech, without either one losing

The problem: a voice and a music bed occupying the same output. Set the music where it sounds good and the words become hard to follow; set it where the words are perfectly clear and the music is so quiet it may as well not be there.

The instinct is to reach for a ducker. There are three better moves to make first.

Move one: choose music that leaves room

Speech intelligibility lives between roughly 1 and 4kHz. A music track with a busy vocal, a bright lead synth, or dense percussion in that region is competing for the same space, and no amount of level adjustment fixes competition: it only decides who loses.

Music that sits under speech well has: no vocal, sparse midrange, and most of its energy either below 500Hz or above 6kHz. Library music is often tagged “underscore” for exactly this. Choosing the right piece is worth more than every processing decision that follows, and it is free.

Move two: cut a hole rather than turning it down

Put an EQ on the music bus with a 2 to 4dB wide cut centred somewhere between 1.5 and 3kHz. The music loses very little apparent presence (that region is not where its character lives) and the voice gains a clear channel.

This is far less noticeable than a 6dB level drop, because you have taken away only the part that was in the way.

Move three: automate before you compress

Manual level automation on the music track (riding it down under speech and up in the gaps) is what film and broadcast mixers actually do, and it beats any automatic process for one reason: it anticipates.

A human knows the line is about to start and begins the move 200ms early, over half a second, smoothly. A ducker can only react after the fact, and it reacts to every breath, every lip smack, every stray noise, producing a bed that surges and retreats constantly. That pumping is the signature sound of an automated podcast intro.

When a ducker is the right tool

For live work, for long-form conversation, and for anything with too many cues to automate by hand, sidechain ducking is the pragmatic answer. Settings that behave:

• Threshold so it triggers on speech and not on breaths.
• Ratio 2:1 to 4:1. It is a level shift, not a wall.
• Attack 10 to 30ms: fast enough to be out of the way by the first syllable, slow enough not to clip the music’s transients.
• Release 300 to 800ms, which is the setting that matters most. Too fast and the music jumps back between words; a longer release keeps it down through a sentence and lets it return in a real pause.
• Depth limited to 6 to 9dB. Many dedicated duckers have a maximum-reduction control; use it. A ducker allowed 20dB of reduction makes the music vanish.
• Filter the sidechain with a high-pass around 200Hz so low-frequency thumps do not trigger it.

How far down should the bed go

A working target: the music sits 15 to 20dB below the dialogue’s typical level while speech is happening. In loudness terms, for a programme delivered at −16 LUFS, a bed under speech commonly measures around −30 to −34 LUFS short-term, rising to −18 or so in the gaps.

Those figures are a starting point. The real test is not a meter: play it on a phone speaker, at a low volume, while doing something else. A bed that is comfortable on monitors at conversational volume is frequently too loud on a phone in a kitchen, because the small speaker has no low end and the bed’s midrange is all that survives.

Sound effects are the same problem

Everything above applies to effects under dialogue, with one addition: an effect that is meant to be noticed should have a gap around it. A door slam landing between two lines reads as loud and clear; the same slam under a line is mush, however loud you make it.

Editing the dialogue to leave the space is more effective than raising the effect, and it is how it is done on any production where the dialogue and effects are mixed by the same person.

Three checks before delivery

1. Mute the music. Does the dialogue sound complete? If it only works with the bed, the dialogue edit has problems the music is hiding.
2. Mute the dialogue. Does the bed make sense on its own, or is it a series of unexplained swells? Pumping is obvious here and hidden under speech.
3. Listen at low volume on a small speaker. Quiet listening is where the balance fails first, because a small speaker and a low level both remove the parts of the music that were not competing.

The free path

• Ardour, Audacity and Reaper all do clip-gain and automation lanes for manual riding: the technique this lesson recommends most, and it needs no plugin.
• Calf Sidechain Compressor and the LSP sidechain compressor are free and have the controls listed above, including sidechain filtering.
• ffmpeg has sidechaincompress for batch work.
• DaVinci Resolve’s free Fairlight page has full sidechain routing.
• Free music with clear licences: the Free Music Archive, ccMixter, and material released under Creative Commons, with the licensing caveats covered in the final module, because “free to download” and “cleared for your use” are different claims.

The limitation

Ducking manages a conflict; it does not resolve one. If the music and the voice want the same frequencies and the same moments, something has to be cut, and the honest answer is usually the arrangement: a shorter music cue, a break under the key line, a different track.

There is also a taste question nobody can settle for you. Broadcast conventions in some countries favour beds so low they are almost subliminal; in others they sit much higher. Complaints about background music drowning speech are a recurring theme in broadcaster correspondence, and they come disproportionately from older listeners and those with hearing loss, whose ability to separate speech from a competing sound is reduced. If your audience includes them, mix the bed lower than sounds right to you.