Exported Video Volume Keeps Changing? Volume Automation, Compressor & Loudness Normalisation Diagnosis Guide

Key takeaways: An exported video whose volume rises and falls between sections is usually an automation, dynamic-processing or export normalisation effect, not a single low fader. Compare the finished file section by section with the timeline, inspect the automation curve, bypass the master-bus compressor and limiter, and check the export normalisation mode before re-exporting a short test.
A finished export plays from start to finish, and no section is silent. Yet one section is clearly loud, the next section is clearly quiet, and a later section is loud again. The change is not a slow drift in one direction, and it is not a single quiet passage in an otherwise even programme. The instinct is often to turn the player volume up, or to push the master fader higher and export again. Neither step addresses the shape, because a single fixed gain change moves every section by the same amount: the loud sections become louder, the quiet sections become louder by the same amount, and the difference between them stays exactly as it was. The finished file still rises and falls, only at a higher level.
The distinction that matters most is between a fixed level and a level that changes over time. A fixed level is one number for the whole programme: the timeline fader, the clip gain, or the player volume. A level that changes over time is a curve, a process, or a setting that treats one section differently from another. Volume automation is a curve. A compressor is a process that reacts differently to loud and quiet material. Loudness normalisation, in one of its forms, treats each section on its own terms. Any one of those can write a rising and falling shape into a finished file, even when the timeline sounded even during editing, because timeline monitoring and export do not always pass through the same stages. This guide separates those layers in order, so that the layer responsible is identified by a change in the finished file, not by guessing.
What shape is this article about?
This article covers only a narrow shape, and several conditions must all hold. First, the subject is a finished exported file: a file that has already been through the export stage. Second, sound is present for the whole duration. No section is completely silent, and the file does not start with sound and then lose it halfway through. Third, the volume rises and falls between sections in a way that can be heard and seen in the waveform. One section is loud, another is quiet, and the change is large enough to notice without straining. Fourth, the shape is the same each time the same file is played. If the loud and quiet sections move to different places on each playback, the shape is not fixed, and the direction points toward playback or encoding behaviour rather than toward the timeline or export settings that are the subject of this guide. Fifth, the sound itself is otherwise normal in pitch and speed, the picture and the sound stay in step, and both channels rise and fall together. If only one channel rises and falls while the other stays steady, that is a channel matter, not this guide's shape.
The reason for drawing the shape this tightly is that three neighbouring guides each state, in their own premises, that fluctuating volume is not their shape. The guide on a quiet export, identified in this library as article 540, states in its premise that the shape it covers is not fluctuating volume, not a gradual fade, and not intermittent sound: its shape is a whole programme that is stably quiet. The guide on a partly silent export, identified as article 522, states in its premise that its shape is not quiet sound, not fluctuating volume, and not clipping: its shape is some sections fully silent while others are fully normal. The guide on a fully silent export, identified as article 537, states in its premise that its shape is not quiet sound, not a gradual fade, and not intermittent sound: its shape is zero sound from start to finish. This article is the mirror image of those three exclusions. It begins exactly where those three guides stop: the file has sound throughout, it is not stably quiet, and its volume rises and falls between sections. That triple premise exclusion is the reason this guide exists as a separate article, and it is also the reason steps taken from those three guides will not diagnose a fluctuating file. A step that measures overall loudness, or that looks for a section that is silent, answers a different question.
Several neighbouring shapes look similar and must not be confused with this one. A file that is stably quiet in every section belongs to article 540, where the fault is in overall loudness or fader level rather than in a shape that changes over time. A file in which some sections are silent while others have sound belongs to article 522, where the silent positions line up with edit points. A file with no sound at all from start to finish belongs to article 537, where the checks are about mute, zeroed faders, and the export audio switch. A file in which only one channel carries sound belongs to article 542, where the fault is in channel mapping rather than in a volume shape that affects both channels together. A file in which the subtitles and the sound do not line up belongs to article 543, which is a subtitle timing layer; this guide does not discuss subtitles. A file in which the sound and the picture hold a fixed offset belongs to article 529, where the sound is late or early by the same amount throughout; in this guide the sound is in step with the picture and only its volume changes. A file whose pitch and speed have both changed belongs to article 504, where the fault is in the sample rate. A recording file that plays normally and then suddenly loses sound halfway through belongs to article 505, which is a recording-stage matter where the first division is a comparison of the two track durations in MediaInfo; this guide is an export-stage matter, the finished file is complete, and the duration comparison from article 505 must not be written as this guide's main test. A recording file with no sound at all belongs to the recording-stage guide, not to an export guide, because the silence was written into the recording before editing began.
The shape also excludes one situation that is easy to mistake for an export fault. If the volume rises and falls only in one player, while the same file plays evenly in a second player, the finished file itself does not carry the fluctuation, and that is a player direction, not this guide's shape.
How it works: the chain from the timeline to the finished file
Sound travels from the timeline to the finished file through several layers, and each layer can change the volume of one section without changing another. The layers are clip gain, track automation, track and bus dynamic processing, the master bus, and the export normalisation mode. Timeline playback passes through the first layers in their monitoring form. Export passes through all of them, and the last layers exist only on the export path, or behave differently there. That is why a timeline can sound even while the export rises and falls: the layers that were even in monitoring were not the layers that wrote the finished file.
1. Volume automation and keyframes are written into the finished file
Volume automation is a curve drawn across time. Instead of one level for a whole track, the level is set point by point, and the editor moves between those points as the playhead travels. In Premiere Pro, the volume of an audio track can be set with keyframes through the Audio Track Mixer or the Effect Controls panel, by showing keyframes and choosing Track Keyframes and then Volume. The Essential Sound panel also carries an Auto-Ducking feature, which creates keyframes on its own: when dialogue is present, it lowers the music or the effects underneath it. An Adobe Community discussion states that keyframes can make audio settings jump back to their initial settings, and it gives the check as expanding the track and looking at Track Keyframes and then Volume to see whether keyframes are present. A guide from caitlinwatson.com states the consequence directly: Premiere Pro renders these adjustments directly into the final video file. Automation is therefore not a monitoring aid that stays behind in the editor. It is written into the finished file, section by section, exactly where the curve rises and falls.
In DaVinci Resolve, the Fairlight page carries full automation lanes with keyframes, while the Edit page allows only the gain of each clip to be adjusted section by section. A guide from Hollyland states the difference directly: Fairlight has full keyframe and automation lanes, and the Edit page has no automation. That difference matters for diagnosis, because a reader who looks only at the Edit page will see no automation curve at all, even when a Fairlight automation lane is quietly shaping the finished file. The check must be made where the automation lives, not where it is easiest to look.
The key distinction inside this layer is between a fixed value and a curve. Clip gain is one fixed number for a whole clip: the entire clip is raised or lowered by the same amount. Track automation is a curve that changes over time: one part of the track is raised, another part is lowered, and the level moves between them. The final volume of any moment is the two added together. Checking only the clip gain will miss the automation layer entirely, because a clip whose gain is correct can still be shaped by an automation curve sitting above it. Conversely, an automation curve that looks flat will not explain a fluctuating file, and the diagnosis must move on to the next layer. The two layers must be read separately and then read together, because only their combination is what the export writes.
A further distinction inside the keyframe layer must be kept clear. The fully silent export guide, article 537, deals with keyframes that set the volume to zero: a keyframe at zero is a silence cause, because the sound at that point is removed. This guide deals with keyframes that set the volume to a non-zero value that changes: a ducking dip, a fade in, a fade out, or a fader movement recorded by hand. Those keyframes do not silence anything. They make one section louder and another section quieter, and that is a fluctuation cause, not a silence cause. The two kinds of keyframe must not be written as if they were the same thing, because they point at different repairs and different guides.
2. A compressor and a limiter are dynamic processes, not fixed gain
A compressor and a limiter are often spoken of as if they were the same device, and as if either one simply made everything louder. They are different devices, they act differently, and neither one is a fixed gain. A compressor narrows the dynamic range: it raises quiet material and lowers loud material, so that the whole programme sounds more even. A limiter sets only a ceiling: it stops peaks from passing a chosen value, so that clipping and distortion are prevented. A guide from Filmora and Wondershare sets the three processes side by side in a comparison table: normalisation adjusts the overall gain toward a target level; compression reduces the gap between loud and quiet; limiting stops peaks from passing an upper limit. The three rows of that table are three different actions, and they must not be swapped for one another in diagnosis.
Fairlight carries a built-in compressor and a loudness meter. Larry Jordan describes a multiband compressor as splitting the sound into four frequency bands, raising the quiet parts in each band while leaving the loud parts untouched, so that the sound is heard as louder while the loudest moment stays unchanged. The same author also gives a clear instruction about where compression belongs: generally, only compress dialogue; never compress music, which is already heavily compressed, or sound effects; and do not compress the master mix — use compressors on dialogue tracks only. That instruction is quoted in full in the source-conflict section below, because it stands in tension with a different source's advice about the master bus, and the two must be presented side by side rather than resolved into one rule.
The practical consequence for diagnosis is that a compressor can create, preserve, or change a fluctuating shape, depending on where it sits and how it is set. A compressor narrows the dynamic range of the material it acts on, and a limiter only stops peaks from passing its ceiling, so the two devices leave different fingerprints, and the bypass test in the diagnostic steps below is designed to separate them.
3. Loudness normalisation comes in two kinds, and they must not be mixed up
Normalisation is a single word for two different operations, and confusing them is one of the most common reasons a fluctuating file is misdiagnosed. Peak normalisation raises the whole programme until its loudest moment reaches a chosen ceiling, for example minus 3 dB or minus 6 dBFS. It is fast, it does not clip, and it pays no attention to how loud the programme sounds to a human ear. Loudness normalisation adjusts the programme according to integrated loudness measured in LUFS, so that the perceived loudness is consistent with a chosen target. The two operations answer different questions: one asks how high the highest peak is, the other asks how loud the programme sounds overall.
Premiere Pro's Audio Gain dialogue carries two peak options, and the difference between them is itself a diagnostic clue. Normalise Max Peak adjusts a group of clips by the same ratio, so the relative sizes of the clips are preserved: a clip that was quieter than its neighbour stays quieter by the same proportion. Normalise All Peaks raises each clip independently to the same ceiling, so clips that were different become the same at their peaks. The first preserves the shape between clips; the second flattens it at the peaks, while still leaving the perceived loudness of each clip free to differ. On the export page, Premiere Pro carries a Loudness Normalisation option, where a target loudness in LUFS and a loudness standard such as ITU BS.1770-3 can be set, according to a guide from Music Radio Creative. In DaVinci Resolve, the Deliver page carries a Normalise Audio option, with presets that include a YouTube preset, according to Larry Jordan. A guide from Hollyland states the limit of peak normalisation directly: peak normalisation cannot guarantee balanced perceived loudness, because clips with the same peak target can still sound uneven when they were recorded in different environments.
The important trap follows from the definition. Peak normalisation is an overall gain change. It raises or lowers the whole programme together, and it does not change the relative size of one section against another. If the finished file is already loud in one section and quiet in another, peak normalisation will push the whole file up or down together, and the fluctuating shape will remain exactly as it was, only louder or quieter overall. That is why the complaint that normalisation was switched on and the file still rises and falls is not a contradiction. It is the expected result of applying an overall gain to a shape that was written by a curve or a dynamic process underneath. The diagnosis must first establish which kind of normalisation was used, and then ask whether the fluctuation follows the timeline automation or follows something in the export layer. Until the kind is named, the word normalisation explains nothing.
4. Per-section loudness comparison is the core measurement of this guide
The quiet-export guide, article 540, compares overall loudness only: the loudness of the whole timeline against the loudness of the whole finished file. That comparison can show that a file is stably quiet, but it cannot show that a file rises and falls, because a single overall number hides the shape inside it. A file that is loud in one half and quiet in the other half can have exactly the same overall loudness as a file that is even throughout. This guide therefore measures section by section. The finished file is divided into sections, for example every 30 seconds, or by scene, and the loudness or waveform amplitude of each section is read in turn, and then the same section is read on the timeline. The comparison is not one number against one number. It is a row of numbers against a row of numbers, and the pattern of the differences is the evidence.
If the shape of the finished file matches the shape of the timeline automation curve — the sections the timeline lowers are the sections the finished file makes quiet, and the sections the timeline raises are the sections the finished file makes loud — then the automation layer is indicated. If the shape of the finished file does not match the timeline — the timeline is flat, and the finished file still rises and falls — then the automation layer is excluded, and the export layer, meaning normalisation, dynamic processing, or encoding, is indicated. The waveform itself is already evidence for this comparison. Opening the finished file in an editor and looking at its waveform section by section shows the rise and fall directly, without relying on the ear to estimate it. The ear is useful for confirming what the waveform shows, and it is not a substitute for the waveform when the shape has to be written down and compared.
| Layer | What it does | How a fluctuation shows up | What excludes it |
|---|---|---|---|
| Clip gain | One fixed number for a whole clip | Every section of that clip moves together | The fluctuation stays when the clip gain is made uniform |
| Track automation | A curve that changes over time | The finished file follows the curve's rises and falls | The curve is flat and the finished file still fluctuates |
| Compressor on a track | Narrows the range inside that track | That track becomes more even, other tracks do not | Bypassing it changes nothing in the finished file |
| Compressor or limiter on the master bus | Acts on the whole mix at once | The finished file rises and falls with the loudest material | Bypassing it removes the fluctuation |
| Peak normalisation at export | One overall gain to a peak ceiling | The whole file moves together, the shape stays | Switching it off leaves the shape unchanged |
| Loudness normalisation at export | Gain toward a LUFS target | The whole file moves toward the target, the shape may stay | Switching it off leaves the shape unchanged |
| Player volume or equaliser | One overall gain at playback | The same file sounds different in different players | The same shape appears in a second player |
Where the settings live in each editor, kept separate
DaVinci Resolve and Premiere Pro put their audio settings in different places, and the two must not be mixed up. In Premiere Pro, track volume keyframes are shown through the Audio Track Mixer or the Effect Controls panel, by choosing to show keyframes and then Track Keyframes and Volume; the Essential Sound panel carries Auto-Ducking and Auto-Match; the Audio Gain dialogue carries the peak normalisation options; and the export page carries the Loudness Normalisation option. In DaVinci Resolve, automation lanes with keyframes live on the Fairlight page, clip gain is adjusted on the Edit page, the Fairlight page carries the built-in compressor and the loudness meter, and the Deliver page carries the Normalise Audio option. The panel names and the exact layout differ between versions, and the descriptions here follow the external guides named at the end of this article; they should not be read as guaranteed to match every version. What does not differ between versions is the logic: the automation must be checked where the automation lives, the dynamic processing must be checked where it is placed, and the normalisation mode must be checked on the export page, because those are three separate layers and a check in one place cannot exclude a fault in another.
Diagnostic steps: the three new tests, in order
The order below matters, because each step narrows the fault to one layer before the next step is attempted. Only one thing is changed at a time, and the finished file is checked again after each change. A short test export is used for every comparison, so that each change can be checked quickly without re-exporting the whole programme for every test. Write every result down before moving on: which sections are loud, which sections are quiet, which setting was read, and what the short test export contained.
Step 0 — Confirm the shape and write down the loud and quiet sections
Confirm that the finished file has sound for its whole duration, and that its volume rises and falls between sections. Listen through the file once, and look at its waveform in an editor, section by section. Write down which sections are loud and which are quiet, in order, with a time for each change: for example, the opening is loud, the middle interview is quiet, and the closing music is loud again. Then play the same file a second time and check whether the loud and quiet sections fall in the same places. A fixed shape points toward the timeline or the export settings, because a setting or a curve writes the same shape every time. A shape that moves between playbacks points toward playback or encoding behaviour, and that direction belongs with the guides on playback and encoding, identified as articles 504 and 522, rather than with the tests below. Also write down the editor and its version, the export preset, and the players used. If any section is completely silent, stop: the shape is not this guide's shape, and the partly silent guide, article 522, must be used first. If the whole file is stably quiet, stop: the quiet-export guide, article 540, must be used first.
Step 1 — Test T1: inspect the timeline volume automation and keyframe shape
This is the first of the three new tests in this guide. Expand every audio track, show the track keyframes, and choose Volume, so that the automation curve for each track is visible. Read the curve section by section, and compare it with the list of loud and quiet sections written down in Step 0. Look for keyframes that set a non-zero value that changes: ducking dips where music is lowered under dialogue, fades in and out, and fader movements recorded by hand. Note the position of each rise and fall in the curve, and ask whether it lines up with the position of each rise and fall in the finished file.
If the positions line up — the sections the timeline lowers are the sections the finished file makes quiet, and the sections the timeline raises are the sections the finished file makes loud — then the automation layer is indicated, and the handling is in that layer: the curve itself must be adjusted, or the Auto-Ducking that created it must be changed, before the next export. If the timeline curve is flat and the finished file still rises and falls, then the automation layer is excluded, and the diagnosis moves to Test T2. A flat curve cannot write a fluctuating file, so a fluctuating file with a flat curve must have been shaped somewhere else.
Two cautions belong to this step. First, check every track, not only the track that carries the dialogue. A ducking curve on the music track will lower the music under the dialogue and raise it again afterwards, and that is a fluctuation in the finished file even when the dialogue track itself is flat. Second, do not confuse this step with the keyframe check in article 537. The keyframes in that guide set the volume to zero and cause silence. The keyframes in this step set the volume to a non-zero value that changes, and cause a rise and fall. A reader who finds a keyframe at zero has found a silence cause, and must follow article 537; a reader who finds a keyframe that dips without reaching zero has found a fluctuation cause, and must stay with this guide.
Step 2 — Test T2: compare loudness section by section, timeline against finished file
This is the second of the three new tests in this guide, and it is the measurement that article 540 does not make. Divide the finished file into sections, for example every 30 seconds, or by scene, and read the loudness or waveform amplitude of each section with a loudness meter: the Fairlight loudness meter, the Premiere Pro loudness meter, or an external meter. Then read the same sections on the timeline, with the same division and the same meter. Write the two rows of numbers side by side, and then write the difference for each section: the finished-file value minus the timeline value, section by section.
If the difference is the same for every section — the whole file is lower or higher by the same amount throughout — then the fault is in an overall gain layer, such as the master fader or the normalisation mode, and that is the layer of article 540 rather than the fluctuation layer of this guide. If the difference changes from section to section — one section differs by a large amount, the next by a small amount, and another by a large amount in the opposite direction — then the fault is in a dynamic or automation layer, and that is this guide's layer. The pattern of the differences, not any single difference, is the evidence. A single section that differs proves only that one section differs. A row of differences that rises and falls proves that the shaping itself changes over time.
This test also separates the two kinds of normalisation. If the difference between the timeline and the finished file is the same in every section, the export normalisation was an overall gain, and the fluctuation in the finished file must already have been present in the timeline shape, because an overall gain cannot create a fluctuation. If the difference changes between sections, the shaping was done by something that treats sections differently, and the diagnosis must move to Test T3, the dynamic-processing layer, or back to Test T1, the automation layer, depending on which shape the differences follow.
Step 3 — Test T3: inspect the master-bus compressor and limiter, then bypass them one by one
This is the third of the three new tests in this guide. Check the master bus and the main output for any compressor, limiter, multiband compressor, or other dynamic effect: in Fairlight, the Dynamics section; in Premiere Pro, any dynamic effect placed on the master or on the export bus; and in the Essential Sound panel, any Dynamics setting. Write down each device found, where it is placed, and what it is set to do. Then bypass them one at a time — one device, one short test export, one check of the finished file — and ask whether the fluctuation changes with the bypass.
If the fluctuation disappears, or changes clearly, when a device is bypassed, then that dynamic-processing layer is indicated, and the handling is in that layer: the device's placement or setting must be adjusted before the next export. If the fluctuation stays exactly the same after every device has been bypassed, then the dynamic-processing layer is excluded, and the diagnosis moves to the export normalisation check below. Bypassing all devices at once proves nothing, because a change in the result cannot then be tied to any one device. The one-at-a-time rule is what makes the bypass a test rather than a guess.
Keep the track-level and master-bus-level devices separate in this step. A compressor placed on a dialogue track only, as Larry Jordan advises, acts on that track alone. A limiter or compressor placed on the master bus acts on the whole mix at once. The two placements must not be mixed up in the written record, because a bypass that removes a track compressor and a bypass that removes a master-bus device answer different questions. The source-conflict section below explains why the two placements are discussed differently by different sources, and why neither placement can be declared correct or incorrect on its own.
Step 4 — Supporting test T4: separate clip gain from track automation
This is a supporting test, not counted among the three new tests, and it is a finer division inside Test T1. For a section that is suspected, read its clip gain first: the single fixed number for that whole clip. Then read the track automation curve above it. Set the clip gain of the suspected section to zero change, or make the clip gains uniform across the sections being compared, and make a short test export. Ask whether the fluctuation is still there, and whether it follows the clip or follows the track.
If the fluctuation follows the clip — the same clip is quiet wherever it is placed, and the quietness moves with the clip — then the clip layer is indicated. If the fluctuation follows the track automation — the quietness stays at the same time position even when the clip gain is made uniform, and it follows the curve's dips — then the track layer is indicated. This step matters because a clip gain that is wrong in one clip can look exactly like an automation dip when only the finished file is heard, and the two can only be separated by making one of them uniform and watching whether the fluctuation follows the change.
Step 5 — Export normalisation mode check: name the mode, then switch it off in a short test
This is a shared layer with article 540, used here only as an exclusion, and it is not counted among the three new tests. Open the export settings and write down the normalisation mode exactly as it stands: whether normalisation is switched on, whether it is peak normalisation or loudness normalisation, which peak option is chosen where a choice exists, and what target loudness in LUFS is set where a loudness target exists. Then switch normalisation off and make a short test export, and ask whether the fluctuation changes.
If the fluctuation stays exactly the same with normalisation off, then the normalisation layer is excluded as the cause of the shape, because a layer that is switched off cannot be writing the shape. If the fluctuation changes, then the normalisation layer is part of the cause, and the handling is in that layer: the mode must be named, and the choice between peak and loudness normalisation must be made deliberately, with the understanding that peak normalisation is an overall gain and will not even out a shape that was written underneath it. This step is the same switch-off comparison used in article 540, and it is used here only to exclude the normalisation layer, not as a new test of its own.
Step 6 — Player comparison and short re-export confirmation
This is a shared layer, used here as a confirmation, and it is not counted among the three new tests. Play the same finished file in VLC and in a second player, and write down whether the loud and quiet sections fall in the same places in both. If the shape is the same in both players, the player layer is excluded, because a player setting cannot write the same shape into two different players. If the shape differs between players, the player layer is indicated, and the file itself is not the layer at fault. Player volume and equaliser settings are overall gain at playback: they raise or lower the whole file together, and they do not change the shape inside it, so turning a player volume up does not repair a fluctuating file.
Then confirm with a short test export. Change only one thing at a time — one automation point, one bypassed device, one normalisation setting, never two or three together — and re-export a short section only. After each change, check the new file's waveform and its per-section loudness, and write the result down next to the change that produced it. Re-exporting with the same settings will produce a file that rises and falls in exactly the same way, so a second export on its own repairs nothing; only a changed setting changes the result. Once a short test export shows the fluctuation reduced or removed in the expected direction, export the full programme with those settings and check the full file once more, because a short test confirms the direction of the change and the full export confirms the result.
| Test | What is changed | What a change in the result means |
|---|---|---|
| T1 automation and keyframe shape | Nothing is changed yet; the curve is read | Finished file follows the curve: automation layer; curve is flat and file still fluctuates: not automation |
| T2 per-section loudness comparison | Nothing is changed yet; loudness is read section by section | Differences are the same in every section: overall gain layer; differences change by section: dynamic or automation layer |
| T3 master-bus bypass | One compressor or limiter bypassed at a time | Fluctuation disappears or changes: that device is indicated; no change: dynamic layer excluded |
| T4 clip gain against automation (supporting) | Clip gain made uniform for the suspected section | Fluctuation follows the clip: clip layer; follows the curve: track layer |
| Export normalisation switch-off (shared) | Normalisation switched off | Shape unchanged: normalisation excluded; shape changes: normalisation is part of the cause |
| Player comparison and short re-export (shared) | Player changed; one setting changed per test export | Same shape in two players: player excluded; result follows the change: that setting is the cause |
The source conflict, presented side by side
Two sources named in the research notes give advice about the master bus that points in different directions, and this guide presents them side by side without choosing one as the conclusion. The reason they differ is that they are talking about different processors and different purposes, and writing them as if they were one rule would mislead the reader.
On one side, Larry Jordan states that the master mix should not be compressed, and that compression belongs on dialogue tracks only. His words, as quoted in the research notes, are: "Generally, only compress dialog. Never compress music (which is already heavily compressed) or sound effects. You should also not compress the master mix. Just use compressors on dialog tracks." The processor in that advice is a compressor. The purpose is to narrow dynamic range inside a track. The placement advised is the dialogue track, not the master mix.
On the other side, moderncreator.app states that a limiter can be placed on the master bus, with its ceiling set at minus 1 dB, and that the input gain can then be pushed by hand to bring the whole mix toward a target of minus 14 LUFS. The processor in that advice is a limiter. The purpose is to prevent clipping while the overall level is raised toward a loudness target. The placement advised is the master bus.
The two statements are therefore not a direct contradiction about the same action. One advises against a compressor on the master mix for the purpose of narrowing dynamic range. The other advises a limiter on the master bus for the purpose of preventing clipping while raising the level. A compressor and a limiter are different devices, as the comparison table in the mechanism section sets out: compression reduces the gap between loud and quiet, limiting stops peaks from passing an upper limit. The purposes are also different: narrowing dynamic range is not the same as preventing clipping. This guide does not resolve the two into a single rule, and it does not write that adding dynamic processing to the master bus is correct or incorrect. The reader must read which processor is being discussed, which purpose it serves, and where it is placed, and then apply the bypass test in Test T3 to the device that is actually present in their own project, rather than following either statement as a general rule.
The loudness target in the moderncreator advice also needs the same care. The figure of minus 14 LUFS is that source's stated target for YouTube, and it is reported here as that source's claim. It is not a standard verified by this site, and different platforms have different targets, so no single LUFS figure should be written as a universal target. The same caution applies to any LUFS figure from the other sources named in this guide: the Filmora and Hollyland guides also discuss platform loudness targets, and those figures are reported here as those sources' claims, not as standards verified by this site. A reader who adopts a target should check the target for the platform the file is actually being delivered to, rather than copying a figure from a guide written for a different platform.
| Source | Processor named | Placement named | Purpose named | What the source states |
|---|---|---|---|---|
| Larry Jordan | Compressor, including multiband compressor | Dialogue tracks only; not the master mix | Narrow dynamic range inside dialogue | Only compress dialogue; do not compress the master mix |
| moderncreator.app | Limiter | Master bus, ceiling at minus 1 dB | Prevent clipping while raising the mix toward a loudness target | A limiter on the master bus, with input gain pushed by hand, can bring the mix to minus 14 LUFS, as that source states it |
Boundaries: what this article does not cover
A file that is stably quiet in every section is article 540. Its premise explicitly excludes fluctuating volume, and its checks are about overall loudness and fader level. A file in which some sections are silent while others have sound is article 522. Its premise explicitly excludes fluctuating volume, and its checks are about edit points and the positions of the silent sections. A file with no sound at all from start to finish is article 537. Its premise explicitly excludes quiet sound, a gradual fade, and intermittent sound, and its checks are about mute, zeroed faders, and the export audio switch. Those three exclusions, stated in the premises of articles 540, 522, and 537, are the boundary of this guide on three sides at once, and they are the reason this guide's tests are per-section rather than overall, and its keyframes are non-zero rather than zero.
A file in which only one channel carries sound is article 542, where the fault is in channel mapping, and both channels in this guide rise and fall together. A file in which the subtitles and the sound do not line up is article 543, which is a subtitle timing layer; this guide does not discuss subtitles. A file in which the sound and the picture hold a fixed offset is article 529, where the sound is late or early by the same amount throughout; in this guide the sound is in step with the picture and only its volume changes. A file whose pitch and speed have both changed is article 504, where the fault is in the sample rate, and the sound in this guide is normal in pitch and speed. A recording file that plays normally and then suddenly loses sound halfway through is article 505, which is a recording-stage matter; this guide is an export-stage matter, the finished file is complete, and only its volume fluctuates. A recording file with no sound at all is a recording-stage matter, not an export matter, because the silence was written into the recording before editing began.
Two further points sit outside this guide even though they feel close to it. The master fader layer of article 540 is a fixed gain, and it cannot explain a shape that changes over time; pushing the master fader higher moves every section together, and the fluctuation remains. The duration comparison of article 505 is a recording-stage test, and it must not be written as this guide's main test, because a complete finished file has no missing duration to compare.
Common misconceptions: seven claims that mislead
- "Normalisation was switched on, so the volume must be even." Wrong. Peak normalisation is an overall gain: it raises or lowers the whole programme together and does not change the relative size of one section against another. A file that was loud in one section and quiet in another stays loud in one section and quiet in another, only at a different overall level. Hollyland states that peak normalisation cannot guarantee balanced perceived loudness, because clips with the same peak target can still sound uneven when they were recorded in different environments.
- "Pushing the master fader higher will repair it." Wrong. The master fader is a fixed gain, and the whole programme moves together when it is pushed. The loud sections become louder, the quiet sections become louder by the same amount, and the difference between them stays exactly as it was. The master fader layer belongs to article 540, and it is not this guide's fluctuation layer.
- "A compressor makes everything the same loudness." Wrong. A compressor narrows the dynamic range, but it does not erase an automation curve: if the curve lowers a section, the compressor acts on the lowered signal, and the section stays lower than its neighbours. A limiter only sets a ceiling, and it does nothing at all to quiet sections that never reach it. The two devices must not be swapped for one another.
- "The timeline sounded even, so the finished file must be even." Wrong. Export has its own normalisation and dynamic-processing layers, which timeline monitoring does not pass through in the same way. Timeline monitoring proves only that the monitoring path was even. It proves nothing about the export path, because the two paths diverge at exactly the points where this guide's faults live.
- "Keyframes only make sound disappear." Wrong. The keyframes in article 537 set the volume to zero, and they cause silence. Non-zero keyframes — ducking dips, fades in and out, and fader movements recorded by hand — only make the volume rise and fall, and they do not silence anything. The two kinds of keyframe must be kept apart.
- "Adding a compressor to the master mix is always correct." Wrong. Larry Jordan states that the master mix should not be compressed, and that compression belongs on dialogue tracks only. Moderncreator.app states that a limiter on the master bus can bring a mix toward a loudness target. The two sources discuss different processors for different purposes, and neither statement should be written as a general rule. The bypass test in Test T3 is the way to decide what the device actually present in a project is doing.
- "Turning the player volume up will solve it." Wrong. Player volume is an overall gain at playback, and it raises every section together without changing the shape inside the file. The player comparison in Step 6 is an exclusion test, not a repair: if the same shape appears in a second player, the file itself carries the fluctuation, and the handling is in the timeline or export layers above.
FAQ
My exported file is loud in some sections and quiet in others, but the timeline sounded even. Where should I start?
Start with Test T1: expand every audio track, show the track keyframes, choose Volume, and read the automation curve section by section against the loud and quiet sections in the finished file. If the curve's rises and falls line up with the finished file's rises and falls, the automation layer is indicated. If the curve is flat and the finished file still rises and falls, the automation layer is excluded, and the diagnosis moves to the per-section loudness comparison in Test T2 and then to the master-bus bypass in Test T3. Do not start by pushing the master fader higher, because a fixed gain moves every section together and leaves the fluctuation unchanged.
How do I tell whether the fluctuation follows the automation curve or the export layer?
Use Test T2: divide the finished file into sections, read the loudness of each section with a loudness meter, and read the same sections on the timeline. If the finished file's shape matches the timeline automation curve, the automation layer is indicated. If the timeline is flat and the finished file still rises and falls, the export layer is indicated, meaning normalisation, dynamic processing, or encoding. The waveform itself is evidence: open the finished file in an editor and look at its amplitude section by section, rather than relying on the ear alone.
I switched normalisation on at export, but the file still rises and falls. Does that mean normalisation failed?
No, it means the kind of normalisation matters. Peak normalisation is an overall gain: it pushes the whole file up or down together, and the fluctuating shape stays exactly as it was. Loudness normalisation adjusts toward a LUFS target, and it also does not erase a shape written by an automation curve or a dynamic process underneath it. Name the mode first — peak or loudness, and which peak option was chosen — then switch normalisation off in a short test export and see whether the shape changes. If the shape stays the same with normalisation off, normalisation is excluded as the cause of the shape.
What is the difference between clip gain and track automation, and why does it matter here?
Clip gain is one fixed number for a whole clip: the entire clip moves together. Track automation is a curve that changes over time: one part of the track is raised and another part is lowered. The final volume is the two added together. It matters because checking only the clip gain will miss the automation layer entirely, and a clip whose gain is correct can still be shaped by an automation curve sitting above it. Use the supporting test in Step 4: make the clip gains uniform for the suspected section and see whether the fluctuation follows the clip or follows the track curve.
Should I put a compressor on the master bus to even out the volume?
There is no single answer, and two sources named in this guide give different advice because they discuss different processors for different purposes. Larry Jordan states that the master mix should not be compressed, and that compression belongs on dialogue tracks only. Moderncreator.app states that a limiter on the master bus, with its ceiling set at minus 1 dB, can bring a mix toward a loudness target, as that source states it. This guide presents the two side by side without choosing one as the conclusion. The way to decide for a particular project is Test T3: bypass each master-bus device one at a time, make a short test export after each bypass, and see whether the fluctuation follows the change.
The volume rises and falls in VLC, but is that a player problem?
Play the same file in a second player and compare the positions of the loud and quiet sections. If the shape is the same in both players, the player layer is excluded, because a player setting cannot write the same shape into two different players. If the shape differs between players, the player layer is indicated. Player volume is an overall gain and does not change the shape inside the file, so turning the player volume up does not repair a fluctuating file.
My file is stably quiet in every section. Is this the right guide?
No. A file that is stably quiet in every section belongs to article 540, whose premise explicitly excludes fluctuating volume. A file in which some sections are completely silent belongs to article 522, whose premise explicitly excludes fluctuating volume. A file with no sound at all belongs to article 537. This guide covers only a file that has sound throughout and whose volume rises and falls between sections, and its tests are per-section measurements, non-zero keyframes, and master-bus bypasses, which are not the tests of those three neighbouring guides.
If the fault still cannot be divided
Write down which sections of the finished file are loud and which are quiet, in order, with a time for each change; whether the timeline automation curve is flat or follows the same shape; the per-section loudness readings for the timeline and the finished file side by side; which compressor or limiter is placed on the master bus, and what happened when each one was bypassed; the export normalisation mode exactly as it stands; and the result of each short test export, and bring that written record with a copy of the finished file and the project file if it is available. The exact panel names for the automation, dynamics, and normalisation settings differ between versions, and the descriptions in this guide follow the external guides named in the research notes; they should not be read as guaranteed to match every version. Bring the written record described above to Room 1209, Dali Plaza, Shek Mun for inspection. The telephone number is 2186 8010 and WhatsApp is 6558 6806. There is no Mong Kok shop and no on-site service. The site's published three-tier fee is Tier 1 HK$400–780, Tier 2 HK$480–1,380, and Tier 3 HK$680–2,800, an overall range of HK$400–2,800, with the actual price subject to post-inspection quote.
Sources
- caitlinwatson.com, "How do I export projects with automated audio levels in Premiere Pro?" — Essential Sound Auto-Ducking creates keyframes automatically; Premiere Pro renders these adjustments directly into the final video file.
- Adobe Community, "Audio settings keep returning to the initial settings" — keyframes can make audio settings return to their initial settings; the check is Show Keyframes, then Track Keyframes, then Volume.
- Hollyland, "How to Edit Audio in DaVinci Resolve: A Step-by-Step Guide" — Fairlight carries full keyframe and automation lanes; the Edit page has no automation; Fairlight carries a built-in compressor and a loudness meter.
- moderncreator.app, "DaVinci Resolve Audio Tips for YouTube Creators" — a limiter on the master bus, with its ceiling at minus 1 dB, and input gain pushed by hand, can bring a mix to minus 14 LUFS, as that source states it.
- Larry Jordan, "Six Audio Techniques To Better Audio in DaVinci Resolve 20" — the Deliver page carries Normalise Audio with a YouTube preset; a multiband compressor raises quiet parts and leaves loud parts untouched; only compress dialogue, do not compress the master mix.
- Wondershare Filmora, "How to Fix Uneven Volume with Audio Normalisation in Premiere Pro" — comparison table for normalisation, compression, and limiting; Normalise Max Peak against Normalise All Peaks; platform loudness targets as that source states them.
- Hollyland, "How to Normalise Audio in Premiere?" — Audio Gain peak options; Essential Sound Auto-Match; peak normalisation cannot guarantee balanced perceived loudness.
- Music Radio Creative, "Audio Loudness in Premiere Pro: How To Fix Volume Issues" — the export page carries Loudness Normalisation, with a target loudness in LUFS and the ITU BS.1770-3 standard, as that source states them.
- Hollyland, "How to Normalise Audio in DaVinci Resolve (3 Methods Explained)" — peak normalisation against loudness normalisation; the Edit page Normalise Audio Levels; the Fairlight LUFS process.