Nobody agrees what a stem is

The word stem means at least three incompatible things depending on who is saying it, and the gap between those meanings is almost always discovered at the worst possible moment — usually on a deadline, usually by someone who cannot fix it themselves.

This is not a knowledge problem. Everyone involved knows exactly what they mean. The problem is that nobody says it, because each party assumes their meaning is the obvious one.

The three meanings

To a mix engineer, stems are grouped submixes of the mix they just finished. Drums, bass, guitars, keys, vocals — five to eight files, each carrying the full mix treatment for that group, every file starting at the same timecode and running the same length. Summed at unity, they reconstruct the mix exactly.

To a mastering engineer, stems mean something narrower: usually four to six groups, delivered with the mix bus processing removed or bypassed, because the entire point is to have access to the balance before the glue went on. A mastering engineer who receives stems with the bus compressor printed on each one has received something that is useless for the job they were asked to do.

To a sync, film or games department, stems mean whatever the delivery specification says they mean, and the specification is the only authority. Often that is a specific set — dialogue, music, effects — with named channel layouts, specific loudness targets per stem, and file naming that will be rejected by an automated system if it is wrong by one character.

None of these is wrong. What is wrong is assuming.

Where it goes wrong in practice

The common failure looks like this. An artist asks their mix engineer for stems, because someone told them they should always have stems. The mix engineer, reasonably, prints grouped submixes with the full mix on them.

Six months later the artist gets a sync placement that needs an instrumental and a TV mix. They send the stems. The music supervisor opens them and finds that the vocal stem has the mix bus compressor's reaction to the vocal baked into every other stem, so muting the vocal does not produce a clean instrumental — it produces an instrumental that breathes in the places the vocal used to be.

Nobody made a mistake. The mix engineer did exactly what was asked. The request just did not contain the information that would have made the answer different.

What to actually specify

A stem request that cannot be misread contains five things. It takes about thirty seconds to write and saves, on average, one re-export and one uncomfortable email.

• How many groups, and which. Not 'the usual' — name them. Drums, bass, gtrs, keys, lead vox, bv, fx.
• Mix bus processing on or off. This is the single most consequential line and the one most often left out.
• Whether they sum to the mix. If they must reconstruct the approved mix at unity, say so, because it constrains how the engineer prints them.
• Start point and length. Everything from the same zero, same length, including tails. Nobody should be aligning files by ear.
• Format and naming. Bit depth, sample rate, and the exact naming pattern.

The second item deserves its own sentence, because it is the one that generates the re-export. Mix bus processing on means the stems sound like the record and cannot be rebalanced. Mix bus processing off means they can be rebalanced and will not sound like the record until they are summed and processed. Both are legitimate deliverables. They are not interchangeable and you cannot convert one into the other after the fact.

The naming problem

Naming seems like the trivial part and causes a surprising share of the pain, because it is the part that gets handled by software rather than people.

A good naming scheme sorts correctly in a file browser, survives being emailed, and tells you what a file is without opening it. That means: a fixed prefix, zero-padded numbers, no spaces if you can avoid them, and the most significant information first.

Something like ARTIST_TRACK_STEM-01_DRUMS_24-48.wav is ugly and completely unambiguous. Drums final FINAL v2.wav is friendly and will ruin someone's afternoon.

The file name is the only documentation that always travels with the file.

Zero-padding matters more than people expect. Without it, a folder sorts stem 10 immediately after stem 1 and before stem 2, which means the order in the browser does not match the order in the session, which means somebody imports them in the wrong order and does not notice until the balance is wrong.

Tails, and the mistake everyone makes once

Print past the end. The reverb tail, the delay repeats, the cymbal decay — all of it needs to be inside the file, which means the print range extends several seconds beyond the last note.

The classic version of this mistake is printing to the end of the audio rather than the end of the sound. The last chord is still ringing at the boundary, so every stem ends with an abrupt cut, and the person receiving them discovers it when they solo something.

Related: print from absolute zero, not from the first note. If one stem starts at the first downbeat and another starts two bars earlier because of a pickup, they will not line up when dropped into a new session, and somebody will spend twenty minutes proving that to themselves.

A default worth adopting

If you want one policy that covers most situations, this is a reasonable one.

Deliver two sets. Set one: grouped stems with the full mix treatment, summing to the approved mix at unity, for anyone who needs to work with the record as it sounds. Set two: the same groups with mix bus processing bypassed, for mastering and for future rebalancing. Both from zero, both the same length with tails, both twenty-four-bit at the session rate, both named identically apart from a suffix.

It is roughly ten extra minutes at the end of a mix. It eliminates the majority of the stem conversations you will otherwise have over the following two years, including the ones that arrive when you have moved on and no longer have the session open.

That last part is the actual argument. The cost of getting stems right is paid once, by you, while the session is open. The cost of getting them wrong is paid repeatedly, at random intervals, usually when recalling the session is expensive or impossible.