How to A/B Two AI Music Tracks Under a Voice-Over Without Fooling Yourself
Two generated tracks, one narration, and a client waiting for a decision. Most editors play both once, pick the one that felt better in the first ten seconds, and then spend the rest of the edit defending a choice they cannot explain. The uncomfortable part is that the first listen is usually the least reliable evidence available: the louder file wins, the newer file wins, the file whose name contains the vendor you already pay for wins. This article sets out a listening protocol that removes those three advantages, costs about half an hour per pair, and leaves a written record that survives a client asking “why this one?”. Every tool fact below was read from SOUNDRAW, Mubert and Loudly’s own licence pages on Last verified: 2026-11-03; measurement references are the ITU-R recommendations named in the sources. Nothing here comes from hands-on testing of the three generators, so no score is given, and anything the published pages do not settle is marked unverified.
If your problem is the mix rather than the choice — ducking, keyframes, delivery levels — that is a different job and it is covered in How to Mix AI Music Under Voiceovers. This protocol starts after the mix is technically finished, when two finished versions both work and only one can ship.
Before you listen, write down what “better” means
A comparison without a fixed criterion is not a test, it is a preference poll, and preference polls change with the weather. Ten minutes of writing before the first playback is what makes the rest of the protocol worth running. Three criteria, in this order, are enough for voice-over work:
- Intelligibility (must-pass). Every word of the narration stays clear at the delivered level, on the device the audience will use. A track that fails this is not a candidate, no matter how good it sounds alone.
- Fit (scored). The music supports the emotional turn of the piece without announcing itself. Judge at the moment the piece changes direction, not at the opening.
- Operational cost (scored). Can you licence it, keep it licensed, edit it, and re-export it without another conversation? This criterion exists because a track you cannot hold at release date is not a cheaper option, it is a future reshoot.
Write the pass line before you press play: “I will choose A if it keeps every word clear and the middle section still lands; otherwise B.” A pre-written rule is the only defence against the very human habit of inventing a criterion after hearing which file you liked. It also gives the client something concrete to disagree with, which is far more productive than arguing about taste.
Two candidates at a time. Not five. Five options make the decision slower and the outcome no better, and each additional file adds another licensing check you will skip.
Step 1: gate both candidates on licence fit first
Listening time is the expensive resource here, so spend the cheap check first. A pair that cannot both be licensed for the actual placement should never reach the speakers — and this is where the three vendors diverge in ways that decide tests before they start.
SOUNDRAW grants personal and commercial use across projects on all plans — the page lists YouTube, UGC, product videos, tutorials, live streams, social media, video games, up to one physical store, client work, hold music, radio and TV. But the same page carries a retention condition that changes what “choose the better track” means: if you use tracks as downloaded, you may keep the content published only while the subscription is active. Two tracks can be equally good and unequally holdable. The page also lists content types that require asking first — video not owned by the user, and theatre or other synchronised content — and three outright prohibitions: registering in Content ID, distributing without modifying, and selling a modified beat on stock platforms.
Mubert Render issues a licence certificate with every download, which is the single most useful artefact in a client handoff. Its page allows monetised YouTube, TikTok and Instagram videos, ads, brand content and podcasts, and client or agency work on the Business plan. Every plan forbids Content ID registration, standalone release on streaming services, stock-site resale and redistribution of the tracks themselves. The detail that matters for a listening test is the duration cap: paid plans are listed at a 25-minute maximum, so a long programme may be assembled from more than one render, and a stitched track is not tested the same way as a continuous one.
Loudly‘s licence agreement (version 19.01.2026) grants paid subscribers a worldwide, non-exclusive, royalty-free commercial licence to synchronise output into audio-only, audio-visual, video, podcast, gaming and software-application projects under 5.1(b)(i). Clause 5.1(b)(ii) then attaches a time condition: projects can be monetised in perpetuity only if they were released during the term of the agreement. Clause 5.2(i) excludes physical products, outdoor, TV broadcast, IPTV, VOD, cablecast, satellite, cinema and theatrical release, and commercial streaming services such as Netflix, iTunes and Hulu. Advertising is available on the Pro plan under 5.1(b)(v), excluding outdoor, VOD, OTT, TV broadcast and theatrical, with a paid-media budget ceiling of 100,000 USD per year.
Read against a real project, those three clauses produce a shortlist rule: if the content will still be online after the plan lapses, SOUNDRAW’s as-downloaded condition and Loudly’s release-during-term condition have to be checked against the go-live date before either track is tested. The general method for reading the rest of an agreement is in the clause-by-clause guide; here you only need the one question “does this track survive to the release date?”.
Whether any of the three tools offers a built-in blind comparison feature is unverified — no interface was exercised for this article — so the protocol below assumes you are comparing exported files, which is the harder case anyway.
Step 2: build two versions that differ in exactly one thing
A valid A/B pair is a controlled experiment, and the control is mechanical rather than clever. Both versions need the same picture (picture lock, not a rough cut), the same narration take at the same performance, the same edit, the same export settings, the same codec and the same upload path. Only the music changes. If the picture moved between versions, the viewer is reacting to the cut, and no amount of listening discipline will separate the two effects afterwards.
Three practical rules keep the pair honest:
- Align the music’s first entry to the same timecode. A track that starts 400 ms earlier changes the perceived pacing of the opening and will read as “more energetic” for reasons that have nothing to do with its composition.
- Keep the music edit identical. If version A needed a 2-second trim to fit its structure and version B did not, you are comparing two edits as well as two tracks.
- Test at the length you will ship. A 20-second excerpt flatters tracks that have no development. If the delivery is six minutes, the test has to include the part of the piece where the music would normally run out of ideas.
Where a vendor’s render cap forces stitching — Mubert’s 25-minute paid-plan maximum is the clear case — build both versions with the same assembly method. Comparing a seamless single render against a stitched one measures your crossfade, not the music.
Step 3: match loudness before anyone hears them
This is the step that decides most A/B tests, and it is the one most often skipped. Human listeners reliably prefer the louder of two otherwise identical presentations; a difference of one or two LU is enough to produce a confident, articulate, entirely wrong preference. The fix is measurement, not willpower.
ITU-R BS.1770-5 is the recommendation behind the loudness and true-peak algorithms used by most modern meters; it defines how programme loudness is measured and how true peak is estimated. In practice you need three numbers per version from the same meter: integrated loudness in LUFS, loudness range, and true peak in dBTP. Then:
- Match the programme, not the music stem. The viewer hears the finished programme, so both versions should land on the same integrated loudness target — the one in the delivery spec, or a single target you choose and apply to both if no spec exists. Matching music stems instead leaves the programmes at different loudness, and the louder programme wins the test.
- Record the music-stem difference separately. A version with a hotter music bed under the same narration tells you something real about how hard the mix is leaning on the music, but it is diagnostic information, not a tie-breaker.
- Never match by peak. Peak level says nothing about perceived loudness, and two files with identical true peaks can sit 6 LU apart.
- Leave headroom. Check true peak after the loudness change, not before; normalising up is the usual way a clean mix acquires inter-sample overs.
Specific platform loudness targets are deliberately not quoted here — they change, and the number that matters is the one in your own delivery spec. Treat any figure you cannot find on the platform’s current documentation as unverified and measure your own export instead.
Step 4: choose three windows that actually carry the decision
A full-programme listen is necessary but inefficient: attention drifts, and the parts that decide the choice are usually three short stretches. Fix them before playback so both versions are judged on identical material:
- The opening 20 seconds. Where a track announces itself. Note the first moment the music draws attention away from the words.
- The densest speech run. Find the stretch where narration continues without a breath — usually mid-piece. This is where masking shows up, and where a track that sounded spacious in isolation turns muddy.
- The turn. The emotional or structural change of the piece. A track that fits the first minute and fights the turn is the most common failure mode in generated music, because the track was chosen for its opening.
Judging order matters as much as the windows. Judge under narration first. A music-only listen is a legitimate diagnostic — it tells you whether the mix is hiding a weak arrangement — but do it last, after the under-narration verdict is written down. Once you have heard a track alone and liked it, the under-narration judgement is contaminated, and you will rationalise the intelligibility problem as “character”.
Step 5: blind the order and hide the identity
Blinding is not bureaucracy; it is the only way to stop the price tag from voting. Four leaks have to be closed, and three of them are easy to forget:
- File names. Rename both to neutral labels — v1 and v2 — and have someone else, or a coin flip recorded on paper before playback, decide which track takes which label. Write the mapping down and keep it out of sight until the log is complete.
- Metadata. Generated files often carry the vendor name in the title or artist tag. Strip the tags, or the blind is broken by the first media player that shows a filename banner.
- Visual giveaways. A waveform or spectrogram shows arrangement density and instrumentation before a single note plays. Compare from a player that shows a plain timeline, or cover the display.
- Order. The first file heard has an advantage. Judge each window in both orders — A then B, and B then A — and count a preference only if it survives the swap.
ITU-R BS.1116-3 describes the formal version of this: blind, repeated presentations of the same material, with the reference hidden among the stimuli. You do not need a listening room to borrow the principle, but you do need repetition and a hidden reference. For a two-track comparison the practical version is the odd-one-out check: play A, then B, then a third file that is a copy of one of them at random, and ask the listener which of the first two it matches. If the answer is at chance across several trials, the two tracks are below the threshold where a preference is meaningful, and the decision should move to operational grounds. ITU-R BS.1534-3 sets out the multi-stimulus variant used when differences are smaller and finer discrimination is needed; it is worth knowing that these methods exist, because “I could hear the difference” is a claim that can be tested.
Step 6: how many passes a decision actually needs
More passes do not produce better decisions. They produce fatigue, and fatigue compresses differences until everything sounds equally acceptable. Four passes, used for different jobs, are the working maximum:
- Pass one — orientation. Whole programme, no pausing, no notes beyond a single line per version. The purpose is to find out whether there is a decision to make at all.
- Pass two — targeted. The three fixed windows, one question per window, both orders. This is the pass that produces the evidence.
- Pass three — discrimination. Only when passes one and two disagree, or when the difference is small. Run the odd-one-out check. If the listener cannot beat chance, stop testing and decide on licence cost and editing speed.
- Pass four — label swap. Rename the files, re-randomise, and see whether the preference follows the music or the label. A preference that flips with the label was never about the music.
Two consistent passes are enough to decide. A third is justified only when a must-pass criterion is in play. A fifth pass in one sitting is not evidence, it is the sunk-cost fallacy with headphones on. If the decision is still open after four passes, the honest conclusion is that the two candidates are equivalent for this piece, and the choice belongs to the criteria you can actually measure: licence exposure, plan cost, and how fast you can re-edit.
What to write down in each pass
The log is what turns a listening session into a decision you can defend. Six or seven lines per version is enough; a page of prose is not. The table below is the version used for voice-over work, with the decision rule attached to each row so that the log cannot be filled in after the fact to fit the answer.
| Log field | How to capture it | Question it answers | Decision rule | What voids the entry |
|---|---|---|---|---|
| Programme integrated loudness | Meter on the final export, both versions, same meter and same settings | Were the two versions equally loud? | Both within 0.5 LU of the same target, or the pass is void | Metering a different export than the one you listened to |
| Music-stem loudness difference | Meter the music files alone, same source render | How hard is the mix leaning on the bed? | Record it, do not decide on it | Comparing a trimmed stem against a full one |
| Word catches | One listener, densest 30 seconds, count words that needed a second listen | Does the music cost intelligibility? | Zero catches to pass; one catch means regenerate or re-mix | Listener reading the script while listening |
| First distraction timecode | Pause at the first moment attention left the words and note the timecode | When does the music start competing? | Later than the opening 20 seconds to pass | A listener who has already seen the piece twice |
| Forced preference plus one reason | Pick one, then write a single sentence naming a moment | Is there a real preference, and is it about the piece? | A reason that cites a timecode counts; an adjective does not | A reason produced after group discussion |
| Confidence, 1 to 5 | Rate immediately, before any re-listen | Is this a decision or a coin flip? | 2 or lower triggers the odd-one-out pass | Rating after the answer is known |
| Licence validity at release date | Plan term against the go-live date, per vendor clause | Can this track still be used when it ships? | Must-pass; a failure removes the candidate | Checking before the final placement changed |
Two log fields are deliberately absent: a ten-point “quality” score and a “would use again” rating. Both invite a verdict before the evidence exists, and neither survives a client asking what the number means.
Four ways this test fools people
Loudness bias. The most common failure and the easiest to fix: the louder file wins on every criterion you thought you were measuring. Step 3 exists because of it, and a test run without loudness matching should be discarded rather than reinterpreted.
Identity leak. The vendor’s name in an ID3 tag, the plan you are paying for, the order in which you generated the two tracks, the file size — any of these can carry the answer. The specific trap for first-time buyers is the price heuristic: the more expensive plan must sound better. It does not, but it will, if you know which file came from it.
The music-only listen. Tracks that sound impressive alone often sit in exactly the frequency range the voice needs. If the only listen was music-only, you have measured the arrangement, not the fit. Under narration first, alone last.
Anchoring on the first eight seconds. The opening is where a generated track is most likely to be interesting and least likely to be tested. A track that wins the first eight seconds and loses the turn is the most common wrong pick in this workflow, and only the fixed windows in Step 4 catch it.
When the honest answer is “neither”
Some pairs should not produce a winner, and forcing one is worse than restarting. Three cases:
- Both fail intelligibility. The problem is density, not identity. Changing vendor will not fix a bed that occupies the speech band; re-brief for sparser arrangement and regenerate. SOUNDRAW’s own FAQ states that every request generates a new list of songs, which makes regeneration the cheap move; Mubert and Loudly both re-render from a prompt without a new purchase.
- Both pass but neither moves you. That usually means the brief was vague, not that the music is bad. Pick on operational grounds — the licence you can hold longest at the lowest cost — and spend the saved time on the edit.
- One fails licence fit. Not a tie-break, an exclusion. Do not run a beautiful test on a track you cannot use; the sunk cost will make you rationalise the licence problem later.
My default recommendation, stated plainly because it is more useful than a balanced non-answer: when both candidates licence-fit and both pass intelligibility, choose the one with the shorter licence exposure and the cheaper plan, and stop. Taste differences under narration are usually smaller than the licence risk, and the time saved is better spent on the cut.
Making the choice defensible to a client
A client who is shown two unbounded options will choose on instinct and change their mind after the export. A client who is shown a decision with evidence will argue once and then move on. Five artefacts are enough:
- Two files, neutrally named, both loudness-matched and both at the delivery target.
- The sealed mapping from the blind test, revealed with the verdict.
- The log, one page, with the must-pass criterion at the top.
- Licence proof — the per-download certificate where the vendor issues one (Mubert does), and a dated capture of the licence clause for vendors whose grant is plan-based rather than per-file.
- One sentence on what would have changed the decision. That single sentence is what stops the conversation reopening three weeks later.
Do not send six versions “so they can hear the range”. That moves the decision to the client without the protocol, which is exactly the outcome the protocol exists to prevent.
Who this protocol suits, and who it does not
It suits: client work where the music choice has to be justified; two candidates that both licence-fit the placement; projects where words must stay clear at the delivered level; teams that want a repeatable house standard instead of an argument per project; first-time buyers who would rather not pay for a plan whose grant ends before the content does.
It does not suit: projects without picture lock, because every timing change invalidates the pair; a single candidate that already passes all three criteria, where testing is theatre; music-led content where the music is the subject and should be judged alone; accessibility-driven decisions, where captions or audio description change the balance and the test has to be re-run; clips under about ten seconds, where there is no turn to test.
Risks and limitations
- Loudness matching removes one bias, not all of them. Two versions at identical integrated loudness can still differ in timbre and in how they mask a specific voice. The protocol controls the largest known confound, nothing more.
- Blind tests produce false negatives. Short exposures and untrained listeners miss differences that become obvious over a full programme, which is why pass two uses fixed windows rather than a single excerpt, and why the odd-one-out check is a trigger for more evidence rather than a verdict.
- One listening condition is not the audience’s condition. A phone speaker and headphones can flip the result. Test on the device the audience actually uses, and log which device it was.
- The protocol costs time. Thirty to forty-five minutes per pair, including exports and metering. On a project with a same-day turnaround that is the whole argument against it, and the honest answer is that a rushed single-pass listen is faster and less reliable, not equivalent.
- Vendor facts are dated. The licence scopes and plan limits quoted here were read on 2026-11-03 and will change; the clause numbers are given so that a reader can check the current text rather than trust this summary.
- No hands-on evidence. No tool was tested for this article, no output was measured, and no score is published. The protocol is a method for producing your own evidence, not a substitute for it.
Related guides
- How to Mix AI Music Under Voiceovers: Loudness, Ducking and Product Demo Handoff — the mixing half of this workflow, including a six-step review pass for a finished mix.
- How to Read an AI Music Licence Before You Buy — clause types and what each one decides for a real project.
- AI Music Licence Checklist for Film Festival Submissions and Distribution — what changes when the finished piece leaves the platform and enters a cinema.
- How to QA AI Music Loops: Clicks, Gaps and Tails — the technical checks that belong before any listening test, because a click will outvote a preference.
- How We Score AI Music Tools — the site’s own evidence standard, and why scores are withheld where no matched test exists.
Official sources and update note
- Loudly — License Agreement, version 19.01.2026 (5.1(b)(i) synchronisation grant; 5.1(b)(ii) in-perpetuity condition tied to release during the term; 5.1(b)(iii) distribution via Loudly’s own service only; 5.1(b)(v) Pro advertising scope and 100,000 USD annual budget ceiling; 5.2(i) exclusions including VOD and commercial streaming; 5.2(iii) Content ID prohibition) — Last verified 2026-11-03
- Mubert — Render licence page and plan table (sync licence, certificate with every download, commercially cleared on Pro and Business, 25-minute duration on paid plans, Content ID and redistribution prohibitions) — Last verified 2026-11-03
- SOUNDRAW — official licensing page (all-plans commercial permission list, Artist-plan distribution condition, retention rule for tracks used as downloaded, “ask first” and “not OK” lists, regeneration behaviour in the FAQ) — Last verified 2026-11-03
- ITU-R BS.1770-5 (11/2023) — Algorithms to measure audio programme loudness and true-peak audio level — Last verified 2026-11-03
- ITU-R BS.1116-3 (02/2015) — Methods for the subjective assessment of small impairments in audio systems (blind, repeated presentations with a hidden reference) — Last verified 2026-11-03
- ITU-R BS.1534-3 (10/2015) — Method for the subjective assessment of intermediate quality level of audio systems — Last verified 2026-11-03
Editor’s note: Score withheld — this article describes a testing protocol and rests on three vendors’ published licence pages plus three ITU-R recommendations, all read on 2026-11-03. No generator was hands-on tested and no output was measured, so no score is given. Volatile facts carry a “Last verified” date; anything the published text does not settle is labelled unverified. Updated 2026-11-03.