Play two songs back to back on a streaming service and they come out at about the same volume, whether one was made in 1975 and one last month. That evenness is not an accident of the recordings. The service measured both, worked out how far each was from a target, and moved a gain control. A few years earlier the opposite would have happened: the newest song would have jumped out of the speakers, because its makers had spent hours making it louder than anyone else's.
Mastering is the craft behind that story. It is the last step before a recording meets the public, done by a person with fresh ears in a carefully built room, and it comes with a single number that every release has to answer for: its loudness. This piece covers what mastering is for, how loudness is measured the way ears hear it, how the loudness war ended and why, the peaks that hide between the samples, and the hearing budget that everything sits on.
- Who takes it
- Year 3 of the Sound Engineering major, and Year 4 of the Music Production specialization of Music. Mastering engineers are a small profession, but every producer and every musician who releases music meets its standards.
- What it is for
- Making a finished mix sound its best on every playback system, making an album hold together, and delivering the right file to each format at the right level.
- Comes after
- Mixing & Signal Processing (36/63).
- Leads to
- No later subject in this major lists it as a prerequisite. It connects to Producing Records & the Music Business (36/70), where release formats and metadata meet money.
- What this piece gives you
- You will not be able to master a record after this. You will be able to read a loudness meter, see why a squashed master gets turned down, and know what a true-peak limit is protecting. Easier first reads: Music Technology & Production (35/73), which has a first loudness demo, and How Recording Changed Music (09/10). This piece uses a different demo, with two complete masters, and goes further into the measurement.
What mastering is for
A mix is finished when the balance is right. Mastering begins when someone else, who has not heard the song a hundred times, listens to it on better monitors in a better room and decides what the release needs. That is usually small: a gentle tilt of the tone, a touch of compression to bind the mix, a correction to something that only shows on a different speaker. It also includes everything that happens between songs and between formats.
| Task | What it means |
|---|---|
| Fresh listen | a second opinion on balance, tone and flaws |
| Consistency | making the tracks of an album sit together in tone and level |
| Sequence and gaps | order, fades and the silence between songs |
| Level | setting the loudness and catching the peaks with a limiter |
| Formats | a version for each destination: CD, streaming, vinyl, immersive |
| Metadata | titles, identifying codes and the details a release needs |
Mastering cannot rescue a bad mix: balance problems between instruments belong back in the mixing stage, and engineers will often ask for a revised mix or for the mix split into groups of tracks ("stems"). What it can do is the last few per cent. The rest of this piece is about the measurement that shaped it most.
LUFS: loudness measured the way ears hear it
For decades the only meters were peak meters: how high does the waveform reach? Peak level says little about how loud something sounds. A single drum hit and a long organ note can have the same peak and differ greatly in loudness. The ear responds to the energy averaged over a short time, and it is not equally sensitive at all frequencies: it hears a 3 kHz tone as louder than a 100 Hz tone of the same level.
The international standard ITU-R BS.1770, in force today in its fifth version (2023), measures this. It does four things. It applies a K-weighting filter that copies the ear's frequency response in two stages: a high-pass that cuts the lowest octaves and a shelf that lifts the highs by about 4 dB. It takes the mean square of the filtered signal in 400 ms blocks, overlapping by 75%. It gates the blocks: anything below −70 LUFS is dropped, and so is anything more than 10 LU below the average of the rest, so silence and fade-outs do not drag the figure down. Then it expresses the average in decibels. The result is integrated loudness in LUFS (the same unit as "LKFS" in the standard), where one LU is one decibel.
Move the slider and watch how much the filter adds or takes away.
The curve is computed from the two filter stages of BS.1770 at 48 kHz. The readout assumes the tone is played in both channels at −20 dBFS (decibels below digital full scale).
Without JavaScript: K-weighting is −1.1 dB at 100 Hz, +0.7 dB at 1 kHz and about +4 dB from 4 kHz up. A −20 dBFS tone in both channels reads −21.8 LUFS at 100 Hz, −20.0 at 1 kHz and −16.9 at 8 kHz.
Notice what this means for a mix: a track packed with energy at 3 kHz will read louder than one with the same peaks and a heavy low end. That is how the ear works, and why mastering engineers watch the meter and their ears together. For broadcast the rule that grew from this is the European Broadcasting Union's recommendation R 128, first issued in 2010 and revised in November 2023: programmes are normalised to −23 LUFS (±0.5 LU), with a maximum true peak of −1 dBTP.
The loudness war, and the normalisation that ended it
A compact disc has a hard ceiling at 0 dBFS. Within it, the way to sound louder than the record before yours on the radio or the jukebox was to lower the dynamic range: compress, then limit, then clip, so the average level climbs toward the ceiling. A reasonable instinct, because louder tends to sound better at first, which is why engineers match levels when they compare. The cost was the music: the drum hit that used to jump above the rest, the verse that was meant to be quiet. Metallica's Death Magnetic (2008) became the standard example when fans found the CD distorted, and a version of the same songs released as downloadable content for the game Guitar Hero III showed much more dynamic range and less clipping. Mastering engineer Ian Shepherd started Dynamic Range Day in March 2010 to argue the case; its slogan was "Dynamic music sounds better."
The war faded when the delivery changed. A streaming service measures each track's integrated loudness and applies a gain so that it plays at a target. Spotify's published defaults are −14 LUFS for its normal setting, −11 for "Loud" and −19 for "Quiet", per ITU-R BS.1770 (Spotify for Artists, as of November 2026). Other services also normalise; their targets are reported by engineers rather than always published by the services (for example, about −14 for YouTube and about −16 for Apple Music), so treat those as approximate and check the current figure before relying on it. A track above the target is turned down; whether one below is turned up depends on the service and the track's peaks.
That removes the reward. A track squashed to −8 LUFS and turned down by six decibels is no louder than one mastered at −14, but it has also lost its dynamics and may carry the distortion of the limiter. Hear for yourself. Both masters here are made from the same invented ten-second song, a quiet verse and a fuller chorus, one gently shaped and one pushed hard through a soft clipper.
Sound isn't available in this reader; the charts and numbers still work.
Loudness is computed in the page with the BS.1770 filters and gating; the lines show loudness over one-second windows. The service applies a single gain to the whole track: target minus integrated loudness, and if that would be an increase it is limited so the sample peak stays at or below −1 dBFS. Real services differ in the details, and Spotify's Loud setting adds a limiter, which this demo does not model. Keep your volume low: "as delivered", the crushed master is much louder.
Without JavaScript: the dynamic master measures −14.2 LUFS with a peak of −1.0 dBFS; the crushed one −7.9 LUFS and −0.1 dBFS. At a −14 LUFS target the crushed master is turned down by 6.1 dB and its peak falls to −6.2 dBFS; the dynamic one is left unchanged. The chorus is 4.4 LU louder than the verse in the dynamic master and 3.2 LU in the crushed one.
Look at the numbers under the charts. After normalisation both masters play at the target, so the crushed one gains nothing, and what is left of the war is the difference you can hear: smaller contrast between verse and chorus, and a flatter, more clipped sound. This is the measure engineers call PLR, peak-to-loudness ratio: the peak minus the integrated loudness, about 13 dB for the dynamic master and about 8 dB for the crushed one. The EBU also defines a loudness range (LRA) for how much the loudness moves within a programme. Not everyone calls the war over: some mastering still aims at the loudest result for players that do not normalise, and a few platforms (Bandcamp, for one) do not normalise at all, so a master has to be judged for more than one place.
True peak: the peaks between the samples
A digital file stores samples, not a continuous wave, and the digital-to-analogue converter draws a smooth curve through them. Between two samples that curve can rise higher than either. This is not an error: it is exactly what the sampling theorem (Digital Audio & the DAW, 36/62) says the samples represent. A limiter that stops every sample at 0 dBFS can therefore leave a waveform that goes above full scale in the converter, a true peak above 0 dBTP, and that can distort in the converter, or after the lossy encoding that streaming applies.
The classic case is a high note whose samples land on both sides of each crest. Take a sine at one quarter of the sample rate, sampled 45 degrees off its crest: every sample is 0.707 of the real peak, which looks like −3 dB, yet the wave behind them reaches full height. BS.1770 therefore defines a true-peak measurement that oversamples the signal to find those hidden crests, in units of dBTP. That is why the guidance is for headroom: Spotify advises a true peak at or below −1 dBTP for its lossy formats (−2 dBTP if the track is louder than −14 LUFS), and EBU R 128 sets −1 dBTP for broadcast.
A sine at a quarter of the sample rate, sampled with the phase you choose and scaled so the highest sample is where the slider says. The curve is the ideal reconstruction (windowed-sinc interpolation at 8 times oversampling, as in the true-peak idea; the standard's exact filter differs a little). Dashed lines: full scale, and the −1 dBTP guidance.
Without JavaScript: with the highest sample at −0.1 dBFS and the samples 45 degrees off the crest, the true peak is +2.9 dBTP. At 0 degrees the two agree.
Limiting and dither: the last two processes
A limiter is a compressor with a very high ratio and a very fast attack: its job is to stop peaks from passing a ceiling you set. It is how a master reaches its final level. Push the signal into it and it turns the peaks down, the same principle you met in compression, taken to the edge. Lightly used, it is invisible. Pushed hard, it trades the shape of every drum hit for a few decibels of loudness, which is exactly the trade the demo above prices. A modern limiter works in oversampled form and can be set to a true-peak ceiling.
The last operation is dither. A mix is usually 24-bit; a CD or a 16-bit file has fewer steps. Each extra bit adds about 6 dB of dynamic range, so 16-bit audio spans roughly 96 dB. Cutting the word length simply discards the lowest bits, and the rounding error becomes a distortion that follows the music, most audible in quiet fades. Dither adds a tiny, steady noise before the rounding, so the error becomes a smooth hiss instead; the usual form is triangular (TPDF), spanning one step either side. Noise shaping can move that hiss toward frequencies where the ear is less sensitive. Dither is applied once, as the very last step, and only when the word length is reduced.
Formats: one song, many destinations
Mastering makes a version for each place the music will be heard, and each place has its own limits.
| Format | What the master must fit |
|---|---|
| CD | 16-bit, 44.1 kHz stereo; dither when reducing a 24-bit master |
| Streaming | loudness normalisation to the service's target; headroom of about −1 dBTP for lossy encoding (Spotify's published advice) |
| Vinyl | physical limits: grooves cannot carry the lowest bass at CD levels, the recording curve (RIAA equalisation, standard since the 1950s) boosts the highs and cuts the bass on cutting, and a 12-inch LP holds about 22 minutes a side at 33⅓ rpm, with less fidelity towards the centre |
| Broadcast | −23 LUFS ±0.5 LU and −1 dBTP under EBU R 128 |
| Dolby Atmos music | an object-based mix delivered as a 24-bit, 48 kHz file; Apple's immersive guidance asks for about −18 LKFS integrated and −1 dBTP true peak |
Immersive formats raise a new question: the sound is not a fixed pair of channels but objects placed in space, rendered at the listener's end for speakers or, by modelling the head, for headphones. A stereo master is still made for the same song, because most people still listen in stereo.
Hearing safety: a budget you spend
All of it sits on top of a pair of ears that cannot be repaired. The World Health Organization says that over a billion young adults are at risk of permanent, avoidable hearing loss from unsafe listening (fact sheet, 3 March 2026). Its safe-listening standard with the ITU (WHO–ITU H.870) uses an adult reference of 80 dB(A) for 40 hours a week, and every 3 dB louder halves the time. It is a budget: spend it fast and it is gone. This is general information, not medical advice. If your ears ring after a night out, or you think you hear less than you used to, see a hearing professional.
Allowed time = 40 hours × 2−(level − 80) ÷ 3. It uses A-weighted levels; a phone's volume setting is not a decibel reading, so a sound-level meter app or a device's own listening-level feature is needed to know where you are.
Without JavaScript: 80 dB(A) allows 40 hours a week, 85 dB(A) about 12.6 hours, 90 dB(A) about 4 hours, 100 dB(A) about 24 minutes and 110 dB(A) about 2 minutes.
What a graduate just sees
- Mastering is a fresh listen and a delivery step: small moves for the whole record, never a rescue for a bad mix.
- Loudness is measured with the ear's weighting and an average, not by the peak; LUFS is that average, and a LU is a decibel.
- Gating means silence and quiet tails do not change a track's integrated loudness.
- A service that normalises erases the reward for being loud, so the squashed master keeps only its costs.
- Peak-to-loudness ratio tells you how much punch is left; a PLR of 8 dB is much flatter than one of 13.
- Sample peak and true peak are different: converters and codecs react to the wave, so leave about 1 dB of headroom.
- Dither belongs at the end, once, and only when the bit depth is reduced.
- Every 3 dB of listening level halves the safe time, so a small increase costs more than it feels.
Where it shows up next
- Producing Records & the Music Business (36/70): release files, identifying codes and what each store or service asks for.
- Sound for Film, TV & Games (36/67): broadcast loudness rules (EBU R 128 in Europe, other regional standards elsewhere) use the same BS.1770 meter.
- Live Sound Reinforcement (36/66): the same decibel arithmetic governs how loud a venue may be for the crowd's hearing.
- In the world: the volume jump between a TV programme and its adverts, the "sound check" and "normalise volume" settings in a music app, and the loudness matching in video and podcast apps.
If you want to go deeper
- 📘 Textbook. Bob Katz, Mastering Audio: The Art and the Science (Focal Press, 3rd edition, 2014/2015), the standard mastering text, with chapters on loudness measurement.
- 📘 Studio reference. David Miles Huber and Robert Runstein, Modern Recording Techniques (Routledge), which places mastering in the whole signal chain.
- 🌐 The standards, free to read. The EBU's loudness pages for R 128 and its explanatory documents, and Spotify's loudness normalization page for artists.
- 🌐 Free course. MIT OpenCourseWare 21M.380 Music and Technology: Recording Techniques and Audio Production (Fall 2016), which covers mixing, mastering and digital audio.
Mastering is the last careful listen and the last decision about level. Loudness is now measured the way ears hear it and equalised by the playback service, so the race to be loudest has no prize, and what remains is the music's shape, a peak that leaves headroom, and ears that last.
Try this: open the settings of your music app and find the volume-normalisation option. Play two songs from different decades with it off, then on. Which one changes the most, and does the loud one still sound better at the same volume?
Sources. ITU-R BS.1770-5 (11/2023), "Algorithms to measure audio programme loudness and true-peak audio level" (ITU, listed in force); EBU R 128 (first issued 2010; revised November 2023, version 5.0: −23 LUFS ±0.5 LU, maximum true peak −1 dBTP, gating at −70 LUFS absolute and −10 LU relative; summary on Wikipedia, "EBU R 128"); Spotify for Artists, "Loudness normalization" (−11, −14 and −19 LUFS settings, −1 dB TP advice for lossy formats, −2 dB TP when louder than −14 LUFS; checked 4 November 2026); Wikipedia, "Loudness war" (Death Magnetic, the Guitar Hero III version, Dynamic Range Day 2010, reported platform targets); WHO, "Deafness and hearing loss" fact sheet (3 March 2026) and the WHO–ITU safe-listening standard (ITU-T H.870); Wikipedia, "RIAA equalization" and "Phonograph record" (playing time and groove limits); Dolby Atmos Music delivery playbook (June 2024) and Apple's immersive loudness guidance as reported by Production Expert; Lipshitz, Wannamaker and Vanderkooy, "Quantization and dither: a theoretical survey," JAES 40(5), 1992. The K-weighting filter coefficients are those of BS.1770 at 48 kHz, checked so that a 997 Hz sine at −20 dBFS in both channels reads −20.0 LUFS. Every number in the interactives and the prose was computed in node.
- 📘 Katz, Mastering Audio
- 🌐 EBU loudness pages
- 🌐 MIT OCW 21M.380