AI Video Audio Loudness and Mix QA: A Delivery Workflow

AI video audio passes QA when the message stays intelligible, the mix has a deliberate hierarchy, peaks survive encoding, and the actual delivery file works on the devices viewers use. A loudness reading is part of that decision, not the decision itself. Start with the delivery specification, fix the balance, measure the complete program, encode it, and listen again.

This guide is for the final audio pass after voice, music, ambience, and sound effects are in the timeline. If you still need to create effects, begin with the text-to-sound effects workflow. If the production question is which generation workflow includes audio, use the AI video generator with audio guide.

Loudness QA begins with a specification

Do not start by searching for a universal social-media LUFS target. The correct target is the one required by the client, broadcaster, ad network, platform delivery document, or internal publishing system. If the destination provides no requirement, define a house specification and write it into the job.

At minimum, record:

The EBU R 128 recommendation is an authoritative broadcast reference based on program loudness, loudness range, and maximum true peak. Its core broadcast target is -23 LUFS. EBU guidance for streaming also discusses higher interim program loudness values in some delivery conditions. Neither document means every online clip should be forced to one number. Use a standard to understand measurement, then use the destination's requirement.

Never write “platform normalized” in a QA note unless you tested the platform behavior and can support the claim. Platforms can change processing, and playback context still varies by device, account, region, content type, and app version.

Understand the four readings that matter

Integrated loudness

Integrated loudness summarizes the measured program over its duration. Use it to compare the final master with the delivery target. Measure the complete deliverable from start to finish. Measuring only the loudest section defeats the purpose.

An integrated number can hide a bad mix. A quiet introduction followed by a crushed ending might average to the target while sounding inconsistent. That is why you also watch short-term behavior and listen.

Short-term and momentary loudness

Short-term loudness reveals changes over a few seconds. Momentary loudness responds faster. These readings help locate a suddenly aggressive end card, a whisper that disappears, or a music hit that makes the rest of the video feel small.

Do not make every section identical. Contrast is useful. The question is whether the contrast serves the story and survives the playback environment.

True peak

True peak estimates peaks that can occur between digital samples when audio is reconstructed. A timeline may show sample peaks below zero while the encoded or reconstructed signal overshoots. A true-peak meter and sensible headroom reduce that risk.

Check the exported file, especially after AAC or another lossy encode. If it exceeds the ceiling, reduce the source or limiter output and encode again. Do not merely relabel the file as approved.

Loudness range

Loudness range describes program variation. It is useful for long-form and mixed material, but it is not a target to maximize. A tutorial watched on a phone in a kitchen benefits from controlled variation. A cinematic website film may tolerate more contrast. Decide for the audience and placement.

Build the mix in narrative order

A reliable hierarchy for information-led video is:

1. Dialogue or voiceover carrying the message.
2. Required sync sounds that explain visible actions.
3. Music that supports pace and emotion.
4. Ambience that establishes continuity and place.
5. Decorative effects that add polish.

The order is not a rule for every film. A no-dialogue product reveal may put the product sound first. The value of the hierarchy is that every level decision has a reason.

Set the voice before the music. Edit obvious breaths, clicks, and unusable takes. Match sentence level with clip gain or gentle automation before heavy compression. If one word vanishes, repair that word rather than raising the entire track.

Bring music in under the voice at a low monitoring level. Low-volume monitoring exposes masking because you cannot rely on raw playback power. If the words disappear, first automate the music around the phrase. EQ may create space, but broad, permanent cuts can make the track lifeless when no one is speaking.

Add sync effects next. A click or impact only needs to be as loud as required to read. Decorative transitions should not repeatedly pull attention away from the message. Ambience should disappear when muted but make the edit feel unnaturally empty without it.

Dialogue intelligibility test

Use a repeatable speech check:

• Listen once without looking at captions.
• Reduce monitor volume until the voice is just comfortable.
• Turn away from the screen and write down any word you miss.
• Replay on laptop speakers and a phone.
• Check proper nouns, prices, calls to action, and words under transitions.
• Confirm that sibilance is not painful and plosives do not trigger the limiter.

Captions improve accessibility, but they do not excuse unintelligible audio. Conversely, do not over-process a natural voice into a harsh, fatiguing signal merely to make every breath loud.

AI-generated or synthesized speech deserves extra checks. Listen for abrupt changes in tone, clipped word endings, inconsistent room character, repeated breaths, unnatural stress, and pronunciation errors. Fix the source where possible. A compressor cannot repair a wrong pronunciation.

Music and effects masking

Masking is not just a level problem. A bright synth, dense guitar, and consonants may compete in the same region even when the music meter appears low. Use four tools in this order:

1. Arrangement: choose a less dense passage under important speech.
2. Automation: lower the music only where the voice needs space.
3. Panning and width: keep critical information stable near the center and verify mono.
4. EQ or dynamic processing: reduce only the competing range when required.

For effects, protect the first consonant of a key phrase. A whoosh ending on the first word can make the entire call to action feel late. Move the effect, shorten the tail, or lower it around the word.

Mono compatibility and phase

Phones and platform players do not always reproduce stereo the way your studio setup does. Some viewers use one earbud. Others listen from across a room. Check a mono fold-down.

Listen for elements that disappear, become hollow, or jump in level. Wide ambience and stereo transitions are common risks. Keep dialogue and required action cues stable. If the music loses its hook in mono, choose a different section or repair the stereo processing.

Do not judge mono by muting one channel. Use a proper mono sum in the editor or monitoring path. Muting a channel tests a missing-channel scenario, which is also useful, but it is not the same test.

End cards, logos, and silence

The final seconds often fail because every track reaches its climax at once. A logo hit, music resolve, call to action, and interface sound compete in a tiny window.

Choose the priority. If the spoken call to action matters, let the music resolve around it and place the logo accent after the key word. If the brand sting is fixed, clear room for it. Follow the AI video end-card handoff when the visual and audio ending need separate control.

Check the first and last frame for unwanted clipped tails. A hard cut to digital silence can sound broken. Add a deliberate fade or preserve the intended room tail, while respecting the exact duration.

Measure only after the balance works

Insert the loudness meter where it sees the final mix. Reset it, play the complete program without interruption, and record integrated loudness, loudness range if relevant, and maximum true peak. Then inspect short-term behavior at the loudest and quietest sections.

If the mix is uniformly too quiet or loud but otherwise balanced, adjust overall gain and remeasure. If only one section causes the failure, repair that section. If the limiter is working constantly, return to the mix. A final limiter should catch controlled peaks, not conceal an unstable hierarchy.

Document both the target and actual result:

File: CAMPAIGN_MASTER_30S_STEREO_v07.mp4
Target: house web specification
Measured: [integrated loudness], [maximum true peak]
Meter/version: [tool]
Review date: [date]
Reviewer: [name]
Exceptions: [none or documented reason]

Test the encoded deliverable

The timeline is not the deliverable. Export the actual codec, open it outside the editor, and replay from beginning to end. Check:

• audio starts on time and remains in sync;
• no transient clips after encoding;
• fades and tails survive;
• left and right channels are correct;
• the file is not accidentally mono or dual mono;
• the voice remains clear at low volume;
• metadata and filename identify the right version.

If possible, upload an unlisted or staging copy and inspect that playback. Platform processing may reveal a problem the local master did not. Compare it with the source at matched perceived level, not with one version simply louder.

Real-device listening matrix

Do not “fix” a phone by destroying the full-range master. If a single placement has a special need, create a named derivative and keep the approved master intact.

Common failure modes

Hitting the target but sounding too aggressive

Short, dense clips can hold high short-term loudness while still meeting an integrated value. Reduce unnecessary compression, restore contrast, and lower the sections that feel relentless.

Quiet voice with loud peaks

One plosive or effect may consume headroom. Repair the transient, even dialogue with clip gain, and use controlled compression. Raising the entire mix only makes the peak problem worse.

Music pumps around speech

Overactive ducking makes the music audibly breathe. Slow or reduce the gain change, automate broader phrases, or select a less dense music section.

The uploaded version sounds different

Verify that the correct master was uploaded, then compare the encoded download or playback with the source. Check loudness matching, channel layout, and true peak. Avoid conclusions based on two players set to different volumes.

Captions and voice disagree

Treat the approved script as the source of truth, then correct the wrong element. A mismatch damages comprehension and makes review difficult, especially for names, numbers, and calls to action.

Final mix QA checklist

• The destination and audio specification are written down.
• Dialogue and key information are intelligible without captions.
• Music supports rather than masks the message.
• Required effects land on picture and decorative effects stay subordinate.
• Ambience is continuous without obvious loops or sudden room changes.
• The mono fold-down preserves every essential element.
• Integrated loudness and true peak meet the actual specification.
• The limiter is not hiding a broken balance.
• The encoded delivery file has been measured and listened to.
• Headphones, laptop, phone, one-earbud, and low-volume checks pass.
• The approved values, reviewer, date, and version are documented.

A professional audio pass ends with evidence: a file that sounds intentional, measurements tied to a real specification, and a review record that identifies exactly what was approved. That is more durable than any platform-target shortcut.