AI MUSIC RANK · Updated October 2, 2026

How to Mix AI Music Under Voiceovers: Loudness, Ducking and Product Demo Handoff

AI Music Rank may earn a commission when you purchase through links on this page. This does not change our comparison or the price you pay.

An AI-generated music track can fit the brand brief and still ruin a product demo. The usual failure is not the generator; it is the mix. Music masks consonants, rises during a feature explanation, pumps between short pauses, clips after encoding, or arrives without enough evidence for another editor to reconstruct the project.

Quick take: clean the narration first, set its level consistently, shape the music around speech, then measure the complete programme—not the music file alone. Automatic ducking is a starting pass. Review every fade manually, measure loudness and true peak against the actual delivery specification, and hand off both the editable session and the AI-music licence record.

Six checks before approving the mix

  1. Narration consistency: inspect level jumps, noise, edits, breaths, room tone and intelligibility; retain a clean dialogue stem and processing notes.
  2. Music masking: inspect consonants, low-mid buildup, busy lead instruments and vocal-like elements; retain music-only and dialogue-plus-music review exports.
  3. Ducking behavior: inspect trigger sensitivity, reduction, fade time, recovery and pumping; retain editable keyframes or sidechain settings.
  4. Programme loudness: inspect integrated, short-term and momentary behavior in the full video; retain the meter report and delivery target.
  5. True peak: inspect peak growth after processing or encoding; retain the true-peak reading and final-file check.
  6. Client handoff: confirm another editor can relink media, change narration and verify rights; retain the session, stems, fonts, cue source and licence packet.

Do not start with a universal LUFS number

ITU-R BS.1770 defines algorithms for measuring programme loudness and true-peak level. The current listed version is BS.1770-5. It gives teams a common measurement method; it does not declare one delivery number for every website, ad platform, conference screen, sales deck, or broadcast channel. Review ITU-R BS.1770.

EBU R 128 is a specific broadcast recommendation. Its current page recommends an average programme loudness of -23 LUFS and also uses Loudness Range and Maximum True Peak Level descriptors. That is useful context for a broadcaster, but it should not be copied blindly into a web-video job. Review EBU R 128.

Ask the destination owner for a written specification. If none exists, create an internal target and test it on the devices and environments that matter: laptop speakers, phone, headphones, a meeting-room display, and the conferencing platform used by sales. Keep the measurement result with the final file so a later revision uses the same reference.

Prepare the narration before adding music

Edit the voiceover for mistakes and timing, establish room tone, reduce obvious noise without creating watery artifacts, control distracting breaths, and apply EQ or dynamics only as needed. Match clips spoken by the same narrator before deciding how far the music must move.

A background track cannot fix inconsistent narration. If one sentence is much quieter, a ducker may pull the music down too late or not at all. If narration clips, lowering the final master will not restore it. Deliver a clean dialogue stem before treating the full mix as ready.

Choose AI music that leaves room for words

Generate or select music after marking the product reveal, feature steps, proof point, call to action, and logo. Prefer arrangements without lead vocals or dense lead instruments during technical explanation. Ask for simpler sections beneath speech and reserve the most active section for an opening, transition, or end card.

If the AI tool offers stems or section control, use them to reduce competing elements before reaching for heavy compression. A quieter mix is not always a clearer mix: a dense synth or guitar in the speech range can mask words even when the overall music meter looks modest.

Premiere: generate ducking keyframes, then edit them

Adobe’s current Premiere instructions say to classify clips by audio type, enable Ducking for Music or Ambience, choose what drives the reduction, and adjust sensitivity, duck amount, fade duration, and fade position. Premiere adds an effect and generates editable keyframes. Read Adobe’s automatic ducking guide.

Review the generated automation phrase by phrase. Pull the music down before important words, avoid a jump in every short pause, and let musical transitions breathe when no explanation is happening. Adobe warns that generating the keyframes again overwrites manual keyframe changes, so save a version before recalculating.

Premiere’s Auto-Match feature can make selected clips more consistent and displays the measured LUFS value, but it does not replace delivery measurement of the finished programme. Review Adobe’s loudness Auto-Match instructions.

DaVinci Resolve: use a dialogue-driven music ducker

Blackmagic’s Fairlight page describes a track FX ducker that lowers a target track such as music when selected dialogue tracks are active. It exposes duck level, recovery, and other controls. Review Fairlight’s automatic ducking workflow.

A track-based setup is useful when a product demo has several voice clips or frequent picture revisions. Keep narration on clearly named tracks, music on its own bus, and the unducked music stem available. Listen for recovery that is too fast, which creates pumping, or too slow, which leaves the music unnecessarily flat after speech ends.

A repeatable review pass

  1. Listen without music: confirm every word, edit, breath, and noise decision.
  2. Listen to music alone: mark structural changes and distracting lead elements.
  3. Generate the first ducking pass: use conservative reduction and editable automation.
  4. Review at low volume: if the narration disappears when playback is quiet, the balance is not robust.
  5. Measure the full programme: record integrated loudness, loudness range, and true peak against the named delivery specification.
  6. Encode and recheck: inspect the actual upload or delivery codec for clipping, level change, and speech clarity.
  7. Ask a second person: have someone unfamiliar with the script note every unclear word and distracting music rise.

Package a reconstructable handoff

Deliver the editable project, consolidated dialogue, music-only stem, mixed stem, final master, textless version if required, caption file, codec settings, loudness report, and a short change log. Use filenames that include project, language, aspect ratio, duration, and version.

For the AI music, include the service, track or generation ID, prompt, creation and download date, active plan, invoice or licence certificate, applicable terms URL, and the finished videos in which the music is synchronized. Note Content ID restrictions, client reuse limits, and whether subscription status affects future edits.

Limits and scoring

No automatic mix feature guarantees intelligibility, accessibility, platform compliance, or client approval. Score withheld: Premiere and Fairlight were not tested with the same narration, music stem, loudness meter, codecs, speakers, and blinded listeners, so this guide explains a defensible workflow rather than declaring a winner.

Related guides

Official sources and update note

Last verified October 2, 2026. Recheck the delivery platform specification and the current software behavior before final export.