AI Voiceover Loudness: Match Clips Without Overload

Confirm that loudness is the problem

First listen for distortion, missing words, background noise and changes in delivery. Loudness matching cannot restore a damaged waveform or make an excited performance sound like a calm one.

If a clip is already clipped, use the clipped AI voice repair workflow to decide whether repair or regeneration is appropriate. If the words are clean but the speaker's delivery differs, the voiceover pickups guide addresses that separate problem.

This guide assumes three clean narration clips from the same project. They need a consistent listening level, not a new voice, a different accent or a rewritten script.

Keep untouched source files. Work on copies and record every processing step. Otherwise, a second normalization pass can make it difficult to understand how the current file relates to the original.

Choose the delivery target before processing

Ask for the destination's actual audio specification or the project's approved target. A podcast, training platform, video soundtrack and phone system may have different requirements. Do not use a number from an unrelated platform simply because it appears in a tutorial.

For the worked example below, the project has chosen an integrated loudness target of minus 18 LUFS and a true-peak ceiling of minus 1 dBTP. These are exercise values, not a universal standard or a recommendation for every delivery.

Integrated loudness summarizes the measured program under a defined measurement method. True peak estimates peaks in the reconstructed signal. They answer different questions: how loud the material measures over time, and whether peaks approach or exceed the allowed ceiling.

Record the measurement settings and the portion of each clip being measured. Very short utterances, long pauses and different content can complicate a direct numerical comparison. Use the same method across the set.

Measure the clips without changing them

Open the source clips in an audio editor with suitable loudness and true-peak measurement. Scan them before applying processing. Adobe's Match Loudness documentation describes one application's measurement and matching workflow.

Write down the results beside the file versions. The following numbers are invented for a calculation exercise, not results from a tested AI model.

The required gain is the target loudness minus the measured loudness. For clip A, minus 18 minus minus 22 equals plus 4. For clip B, the same calculation gives minus 2.

That calculation is a starting point for uniform gain. It does not tell you whether the peaks can tolerate the increase, which is why the next check matters.

Calculate the peak consequence

With only a uniform gain change, peaks move by the same number of decibels. Clip A's predicted true peak becomes minus 3 dBTP after adding 4 dB. Clip B's becomes minus 4 dBTP after reducing it by 2 dB.

Clip C is the important exception. Adding 6 dB to its minus 3 dBTP peak predicts plus 3 dBTP. That exceeds the example's minus 1 dBTP ceiling by 4 dB.

You cannot make clip C meet both targets with that gain change alone. A “normalize” button does not remove the conflict. The file needs an appropriate dynamics decision, a better source or an approved change to the delivery plan.

If a processor automatically limits peaks, inspect what it did. Meeting a numeric ceiling while audibly flattening or distorting the narration is not a successful result.

Decide how to handle the difficult clip

Listen to clip C and locate the peaks that restrict the gain. Are they isolated loud syllables, excessive plosives or part of a deliberately dynamic delivery? The answer changes the useful correction.

A careful editor might adjust an isolated event or apply restrained dynamics processing. If the source has an inconsistent performance, regenerating that passage can be cleaner than forcing heavy processing onto it.

In Voice Lab, create a short comparison source using the same approved voice and delivery settings where available. Keep the spoken text free of processing instructions:

The new recording still requires measurement and listening. Voice generation is the source step; the external audio editor performs loudness analysis and matching.

Another option is a lower project target, but only if the delivery requirements allow it and the owner approves the change. Do not quietly lower the entire set to accommodate one poor clip when a fixed specification must be met.

Listen across joins at a fixed playback level

After processing, place the clips in the intended order. Keep your monitoring level fixed and listen through the transitions. A matched integrated value does not guarantee that the opening sentence of every clip feels equally loud.

Check short instructions separately. A two-word response does not provide the same measurement context as a long explanatory paragraph. Compare it with neighboring material and the project's reference rather than chasing tiny numerical differences.

Also listen for changes in tone, distance and background character. Those can be mistaken for loudness differences. More gain will not fix a hollow-sounding pickup, and equalization should not be used automatically without identifying the cause.

Trim unnecessary leading or trailing silence carefully. Preserve intentional pauses and breaths. Editing silence merely to influence a meter can damage the narration's pacing without improving the listening experience.

Measure the delivered file again

Export using the required format, then reopen and measure the exported rendition. Encoding and subsequent processing can affect peaks. Keep the export check separate from the measurement of the working session.

If the narration will be mixed with music, review it in that mix too. A voice that is consistent alone may be masked by a dense music passage. That is a mixing decision, not evidence that every voice clip needs more normalization.

The AI music mastering guide covers broader music delivery considerations. Do not copy a music target into the narration project without checking the actual destination.

Use the voiceover loudness worksheet to record source values, proposed gain, predicted peaks, chosen treatment and final export measurements. The example rows are clearly marked as calculations; fill the observed fields with your own measurements.

Keep the approved reference with the set

Save one accepted reference passage and the project's delivery specification alongside the finished clips. Record the editor and measurement settings used. A future pickup should have something concrete to match.

Retain the source and processed versions with clear names. If a client later changes the target, you can return to the clean master rather than repeatedly processing an already limited file.

Approval should include both the measured requirements and the listening review. The useful outcome is a set that meets its destination's needs and plays naturally in sequence, not merely a table of identical numbers.

Frequently asked questions

Match the set, then trust the playback

Measure first, check the gain calculation and resolve any peak conflict deliberately. Finish by listening through the actual delivery. The numbers should support a consistent narration experience, not become a substitute for it.