Mastering Basics

Mastering is the last set of decisions made on a finished mix, and most of them are about context rather than sound. It asks how this track sits against the others on the record, against whatever the listener plays next, and against every system it lands on — a phone speaker, a car, headphones on a bus. One engineer, one stereo file per song, a short list of moves.

Two things get misunderstood, and both cost money. The first is treating mastering as a repair stage, the place where a mix that never worked finally will. The second is still mastering for loudness. Every major streaming service now sets playback level on your behalf, so the loud master and the quiet one reach the listener at the same volume — and only one of them paid for the trip.

A Master Is a Delivery Decision, Not a Second Mix

A mastering engineer works on a stereo sum, so every tool reaches everything at once. Broad moves are available and often decisive — a decibel out of a wide band around 300 Hz can unclutter a whole record. Surgical ones are not. If the snare is three decibels too loud, nothing done to the two-track fixes it without dragging the guitars along, and the honest answer is a mix note. Buses, sends and automation belong to the mix, and they are already gone.

What mastering does decide is what a single mix session could not see. The level relationship between track four and track five. Whether the record holds one tonal thread front to back or wanders. Where the peak ceiling sits, which file each retailer receives, which identifiers get embedded. None of that is audible on one song played alone — which is exactly why it gets skipped.

Every Major Platform Now Sets Your Playback Level For You

Loudness normalisation measures how loud a track is and applies a fixed gain so everything plays back at roughly the same level. The measurement follows ITU-R BS.1770 and reports in LUFS — loudness units relative to full scale, a unit carrying its reference in its name in exactly the way the decibel lesson insisted a level must. Three readings exist and only one is acted on. Momentary uses a sliding 0.4-second window and no gating; short-term uses a sliding 3-second window. Integrated covers the whole track and is gated, discarding blocks under −70 LUFS and blocks more than 10 LU beneath the track's own gated average. Integrated is what a service reads. The other two are for watching while you work.

Every figure here was checked against the platform's own documentation on 12 August 2026, and nothing a platform does not publish itself appears. Spotify states a target of −14 LUFS integrated and normalises in both directions: negative gain for masters above it, positive gain for masters below. Premium listeners can pick another — Loud at −11 LUFS, Normal at −14, Quiet at −19 — so where your track lands is partly a listener setting. Apple Music normalises through Sound Check, which measures a track, stores the figure in the file's metadata and uses it to raise or lower playback. Apple names ITU-R BS.1770 as a comparable approach but publishes no LUFS target. YouTube publishes none either: its encoding specification for music videos covers codec, sample rate, bit depth and channel count, then stops. The levelling feature it does document, Stable volume, is a different mechanism: it adjusts continuously inside a video rather than setting one gain for it, and it is off for YouTube Music and official music videos.

Numbers circulate for other services that their own documentation does not carry, and that is the trap. Even published behaviour is conditional. Spotify's upward gain is bounded by the track's own headroom and leaves a decibel free for lossy encoding, so a quiet master with peaks already at the top does not get the full lift. On the Loud setting a limiter engages at −1 dB in sample values, 5 ms attack and 100 ms decay — a very dynamic master can end up limited by the platform rather than by you. Album context changes it again: an album is normalised as a unit, so the relationships between its tracks survive, while shuffling normalises each track alone.

The Loudness War Ended and Loud Masters Lost Twice

For decades a louder master was louder on the radio, louder in the shop and louder than the previous track in a CD changer — so records got squashed harder each year to keep up. Normalisation removes the prize. Two masters of the same song delivered at −6 and −14 LUFS integrated reach the listener equally loud, because the service applies eight more decibels of attenuation to the first. Apple says the consequence out loud in its own delivery guidance: tracks mastered loud get played back lower, which can make them sound weaker.

What the loud master paid is peak-to-loudness ratio, the distance between a track's average level and its highest peaks. A brickwall limiter earns integrated loudness by holding peaks down and lifting everything underneath, so that distance is the currency. The AES recommendation for streaming puts it plainly: a recording with a high peak-to-loudness ratio is often perceived as clearer and less fatiguing than one that has been excessively peak-limited. A kick transient pinned flat while the sustain rises around it arrives as weight without impact — and nothing done afterwards puts the front of that note back.

A File That Never Touches 0 dBFS Can Still Clip

A sample-peak meter reports the highest number in the file, which is not the highest level the signal reaches. Samples are points on a curve, and a converter reconstructs the curve running through them — between two samples sitting at −0.3 dBFS, that curve can rise above full scale. Those are inter-sample peaks. True peak is the level after the signal has passed a mathematically ideal converter and reconstruction filter, written in dBTP, and it can exceed 0 dBTP in a file whose sample peaks never do. Records cut before true-peak-aware limiters are full of them.

Measuring it takes oversampling. A BS.1770-compliant reading needs the signal sampled at 192 kHz or higher, so a 44.1 or 48 kHz master gets upsampled inside the meter — and even then a typical meter carries under 0.6 dB of error. That error is one reason the recommended ceiling is not zero. The AES recommendation for streaming is that maximum true peak not exceed −1 dBTP at the input of a lossy encoder. Spotify asks for below −1 dBTP, and below −2 dBTP if the master is louder than −14 LUFS integrated. Apple's advice is to leave at least a decibel of headroom, because the four-times oversampling used in playback conversion will clip material already sitting at 0 dBFS.

The encoder is the other reason. A decoded AAC or MP3 is not sample-identical to what went in, and its peaks can land above the original's — overshoot that grows as the bitrate falls. The AES notes a 256 kbps coder may be satisfied by a limiter threshold as high as −0.5 dBTP, while lower rates can need one below −1.0. So −1 dBTP is a sensible default rather than a guaranteed safe number, and where a service publishes something stricter, use the stricter figure.

An Album Is a Running Order, Not a Folder of Files

Sequencing is a level decision as much as a musical one, and normalisation decides which of your decisions survives. Played as an album, the record is normalised as a whole, so the quiet closer stays quiet against the opener and the arc holds. Played on shuffle or inside a playlist, each track is normalised alone and that arc flattens. Both happen to the same release — so sequence for the record, then audition the tracks alone, because the listener chooses and you do not.

Spacing is the part left to the distributor and regretted later. The silence between two songs carries musical weight — a short gap reads as a breath, a long one as the end of a side — and a segue has to be built rather than hoped for. Decide fades on purpose, trim heads so nothing clicks in, and leave enough tail that the last reverb is not amputated mid-decay.

Then metadata, the least interesting and most consequential part of a delivery. An ISRC is a twelve-character identifier for one specific recording, permanent and independent of the format it appears on. Only one is ever issued to a track, and one can never stand for more than one recording, so an alternate take, a radio edit and a remix each need their own. It identifies the recording rather than the song, the release or the performer — which is why a wrong one misroutes credit and payment for years.

Before You Send It Anywhere

• Read integrated, not momentary — a service applies one gain to the whole track, so a momentary meter flickering near −9 LUFS tells you nothing.
• Set the limiter ceiling in dBTP — −1 dBTP is the streaming recommendation, lower for loud masters or low-bitrate encodes; −0.1 dBFS on a sample-peak meter is not a ceiling.
• Stop aiming at a loudness number — normalisation moves you onto it from either side, so your level choice only controls how much peak-to-loudness ratio you spent.
• Compare versions at matched loudness — an unmatched A/B between a limited and an unlimited master is mostly a level comparison, the bias normalisation exists to remove.
• Sequence for the album, then hear tracks alone — album normalisation preserves the relationships between songs and track normalisation does not, and both will happen.
• Send it back to the mix when it is a mix problem — a stereo file cannot move one instrument, and no amount of mastering buys what the vocal fader would.
• Hire out when the room is the weak link — most of what a mastering engineer sells is a second opinion formed on monitoring you do not own.
• Do it yourself when the stakes are small and the date is real — a single going to a few hundred listeners needs a better mix, not an outside master.