Mixing in the Box: Buses, Sends and Automation

A mix inside a DAW is a routing structure before it is a set of plugin settings. Every track feeds something, that something feeds something else, and where a processor sits in that graph decides most of what it is able to do. The balance lesson in this track owns faders and pan positions; this one owns the wiring underneath them — which signals share a fate, which share a room, and how any of it changes between bar 8 and bar 40.

Two beliefs get in the way, and they pull in opposite directions. The first is that a floating-point mix engine retired gain staging, so levels inside the box stopped mattering. The second is that a mix is a photograph — one good static balance, rendered once. Both are half right, and both produce the same symptom: a session that measures fine and sits completely still.

The Summing Bus Has No Ceiling and Four Things Downstream Still Do

Fixed-point audio counts in evenly spaced steps, which is where the headroom lesson's 6.02 dB per bit comes from. Floating point does not count that way. A 32-bit float sample is one sign bit, eight exponent bits and a 24-bit significand, and the last two do different jobs: the significand fixes precision, the exponent fixes position. Because the exponent moves, quantisation noise stays a fixed proportion of the signal rather than a fixed level — the float noise floor rises and falls with whatever it is carrying.

So do not take 6.02 and multiply it by 32. That rule describes evenly spaced steps, and float steps are not evenly spaced — the 6.02 attaches to the exponent instead, where each increment doubles the value. Your 24-bit significand holds roughly 144 dB of signal-to-noise wherever you happen to be sitting, and the eight exponent bits slide that window across about 254 doublings, on the order of 1,500 dB of absolute range. A bus running 30 dB over full scale is not damaged; pull it back down and the signal returns with the ratio of signal to noise it had up there. Ableton's manual says as much outright: tracks can be driven far into the red without clipping.

Then the number has to become a real, bounded voltage, and there are four places that happens — the master feeding your interface, a hardware insert that hits a converter twice, the bounce if it is a 16- or 24-bit file, and the lossy encode, which can overshoot between samples. That last one is why European broadcast caps true peak at −1 dBTP and Spotify asks for −2 dBTP on masters louder than its −14 LUFS playback target. There is also a fifth consequence that never shows up as clipping: a threshold is an absolute level, so a bus running 15 dB hot hits every compressor and modelled circuit downstream 15 dB harder. Nothing breaks. Everything is different.

A Bus Is Where You Stop Making the Same Decision Eight Times

Grouping is introduced in the balance lesson as a convenience. What it actually buys you is structural: a bus collapses many decisions into one, so eight drum mics become one fader, one compressor instance, one insert point and — the part that matters later — one automation lane. Anything you would otherwise do eight times, badly and slightly differently, you now do once.

Buses sum, and the arithmetic carries over from the headroom lesson. Eight channels peaking at −6 dBFS do not arrive at −6; uncorrelated material stacks toward +3 dB per doubling and correlated material toward +6, so a full kit lands well above any single mic in it. Inside a float engine that costs nothing on the bus itself — but it has moved the input to every threshold sitting on that bus. Set the bus fader before you set the bus compressor.

Build the topology before you process into it, and know the difference between two controls that look identical. An audio bus sums its members and sits downstream of them, so its fader rides after any bus processing; a VCA-style group fader scales its members' own faders and therefore rides before it. Pull the VCA down and the bus compressor hears less. Pull the bus fader down and it hears exactly what it heard. A bus is also a claim that these tracks belong together — two guitars needing opposite EQ moves are two decisions, and grouping them only relocates the argument.

Reverb Belongs on a Send Because a Room Is Something Everyone Is In

An insert sits in line. The whole signal passes through it and the processed version replaces the original, on that channel only, which is exactly what you want from EQ, compression and a high-pass filter — tools whose job is to change one specific sound. A send does something different: it taps a copy at an adjustable level and lets the original carry on, routing the copy to a return where the effect runs fully wet.

Reverb belongs on the second kind because a reverb is a room, and a room is shared. Put an instance on every channel and you have built twelve rooms, which is the same as building none — nothing coheres, because no two instruments are anywhere near each other. One reverb on a return, fed by every channel at its own send amount, puts the whole band in one space at different distances from the listener. The send knob stops being an amount and starts being a distance.

Where you tap the copy decides how it behaves. A post-fader send scales with the channel fader, so the wet-to-dry ratio survives a fader ride — pull the vocal down 3 dB and its reverb follows it down. A pre-fader send ignores the fader entirely, which is what monitor mixes need, and also means a channel you have faded to silence is still feeding the return. Post-fader is the right default for effects for exactly that reason.

Parallel Processing Adds a Copy Instead of Replacing the Original

Parallel compression blends an untouched copy of a signal with a heavily compressed version of the same signal, and the result is not a gentler version of what an insert does — it is the opposite move. An insert compressor pulls the peaks down. A parallel blend leaves the peaks alone in the dry path and raises the quiet detail underneath them, because the crushed copy's soft passages have been brought up close to its loud ones. The transient survives in the dry path, which is why the technique keeps an immediacy that insert compression flattens.

This only exists because you can route a copy somewhere. A wet-dry knob on a plugin is the same structure with the wiring hidden — knowing that is what lets you build the version the plugin does not offer, like a drum bus copy through a filter and a limiter, blended a long way underneath.

Two paths mean two arrival times. If one is delayed relative to the other, the sum is comb filtering rather than reinforcement — the same problem the phase lesson describes when one mic sits farther from the source than another. Look-ahead limiters and linear-phase EQs delay their path by real milliseconds, and modern DAWs compensate for that automatically, including on return tracks. Three situations still bite. Low-latency monitoring modes deliberately drop compensation on input-monitored tracks, which can then sit out of sync with returns that are still compensated. Hardware inserts add converter round-trip latency the host cannot know without being told. And a plugin that misreports its own latency is uncorrectable by definition.

A Mix That Sits Still Is a Mix You Stopped Early

Automation stores a parameter's value against time, which turns any control in the session into a performance rather than a setting — and that is the thing separating a mix from a rendering.

A static balance is one compromise stretched across an entire song, and it cannot be right everywhere. The vocal level that sits perfectly in a sparse verse is buried the moment the guitars arrive. The fill that lands beautifully in bar 8 is 3 dB too loud when it repeats in bar 40 over a fuller arrangement. Engineers who do this for a living ride vocals almost syllable by syllable, and the lane ends up looking like a jagged series of small corrections rather than an elegant curve. They automate send amounts too — throwing reverb onto the last word of a line and pulling it back before the next — and use mutes and filters to build arrangement that was never tracked.

Learn what the write modes differ on, because the names do not tell you. Touch returns to the existing automation when you let go of the control; latch keeps writing whatever you left it at. Trim writes an offset against what is already there instead of replacing it — so you can lift a fully automated vocal by 1 dB without redrawing an hour of moves. Write it all against routing you have finished, because moving a channel into a group afterwards nests its fader moves under a second fader that is also moving.

A Louder Reference Always Wins, and Other Reasons to Stop

• Match your reference by loudness, not by peaks — matching peaks tells you nothing about matching loudness, which is why broadcast abandoned peak normalisation. Match integrated LUFS instead.
• Assume the louder one wins before you start — an industry spent two decades competing to be louder, which only makes sense if louder was being rewarded. Unmatched is uncompared.
• Reference one attribute at a time — low end, vocal position, stereo width. "Does mine sound as good" has no answer you can act on; "is my kick 4 dB bigger" does.
• Judge the export, not the session — the render is the version with fixed-point walls and a lossy encode in front of it. Leave true-peak headroom rather than printing 0.0 dBFS.
• Set the bus fader before the bus processor — summing already moved the level arriving at every threshold on that bus, and no threshold cares that your engine has headroom to spare.
• Null-test any parallel path you build — bypass the processing on both sides, flip polarity on one, listen for silence. Anything you hear instead is a timing offset.
• Stop when your edits stop being fixes — if you are flipping between two settings you cannot tell apart, the mix is finished and your ears are what changed.