Audio is the fastest way to tell whether a video was planned or simply assembled. Viewers forgive imperfect framing, flat color, even a slightly shaky handheld shot. They rarely forgive a soundtrack that fights the edit, narration that drifts out of sync, or loudness that jumps between scenes. Generative audio has collapsed the cost of producing music and voice to nearly zero. It has not collapsed the cost of producing sound that actually fits.
This guide covers a repeatable workflow for AI music and synthetic voiceover in video projects: planning the sound before generating anything, prompting for usable stems, directing a voice performance, syncing to picture, mixing for real playback environments, and running quality control before publishing. It stays deliberately tool-neutral. The same process applies whether you generate inside a browser app, a mobile editor, or a desktop nonlinear editor with a dedicated audio page.
Why Audio Decides Whether a Video Feels Finished
Three failure modes account for most complaints about AI sound in video. The first is wrong energy: a track that is technically impressive but pushes against the pacing of the cut. The second is wrong tone: playful ukulele under a story about loss, or ominous drones under a cheerful product demo. The third is wrong loudness: dialogue at one level, music at another, ambience at a third, with no relationship between them.
Each failure is a planning problem, not a generation problem. The tools will happily hand you something polished. They cannot know that your second act needs restraint, or that your host speaks at 165 words per minute and needs a bed with space between the kicks.
Treat audio as a deliverable with its own review pass. Export the video with the sound muted and play it back. If you can still follow the story, the visuals are carrying their weight. Then listen with the screen off, the way a podcast listener would. If the narration needs the picture to make sense, your script has gaps that no amount of mixing will fill.
Finally, accept that audio quality is judged faster than video quality. Viewers tolerate a soft image for minutes; they wince at a clipped consonant in seconds. Budget review time accordingly.
Plan the Sound Before You Generate a Single Second
The most common workflow mistake is generating music first and editing to it. That inverts the process and forces you to cut around a track you did not choose deliberately. Reverse it: lock the story, then build sound around the locked cut.
Map the timeline in cues, not minutes
Create a cue sheet with timecodes, the purpose of each cue, and a pass/fail criterion. A simple version looks like this:
- 0:00-0:06 — Hook. Purpose: stop the scroll. Criterion: energy present within one second, no slow fade-in.
- 0:06-0:28 — Setup. Purpose: support narration. Criterion: sparse arrangement, no melodic line competing with speech.
- 0:28-0:34 — Turn. Purpose: mark the shift in argument. Criterion: a clear musical event at the cut point.
- 0:34-0:58 — Payoff. Purpose: emotional lift. Criterion: full arrangement, raised energy, no vocal samples.
- 0:58-1:00 — Button. Purpose: resolve. Criterion: clean ending, not a hard cut mid-phrase.
This sheet becomes your prompt plan. Instead of asking for one perfect two-minute track, you generate three or four beds that map to the structure, then blend them at the seams.
Write a five-line sonic brief
Vague prompts produce generic output. A brief forces specificity:
- Emotional arc — restrained curiosity that opens into confidence.
- Reference palette — warm analog textures, brushed drums, soft electric piano.
- Instrumentation — drums, upright bass, Rhodes, light vinyl noise; no brass, no choir.
- Tempo and energy — 92 BPM, medium-low, steady with a lift in the final third.
- Constraints — instrumental only, nothing busy between 1 kHz and 4 kHz, clean tail for an outro.
Keep the brief in a project file. When you return to the project weeks later for a revision, the brief tells you why the track sounds the way it does — and what to change without starting over.
Writing Music Prompts That Actually Fit the Edit
Music generation responds to descriptive language in much the same way image generation does: concrete nouns beat abstract adjectives, and contradictions average out into mush.
Describe instrumentation, genre, and era
"Epic cinematic" is a mood, not a description. "Low strings, taiko hits, a single sustained cello note, 1970s orchestral score aesthetic" is a description. Era cues are especially useful because they carry production conventions with them: tape saturation implies the 1960s and 1970s, gated reverb implies the 1980s, sidechained pads imply modern electronic pop. Those conventions do more work than any list of adjectives.
If you have reference material, use it as a texture guide rather than a copy request. "Similar warmth and room size to a small jazz club recording" communicates a reverb character and a dynamic range without asking the model to imitate a specific artist.
Control tempo, key, and energy curve
Tempo is the single highest-leverage parameter for editing. If your cut points land every four seconds, a track at 120 BPM gives you a beat every half second and a bar every two seconds — mathematically convenient. If your tool accepts BPM, set it. If it accepts a key, choose one that keeps the music out of the fundamental range of the narrator's voice; a bed in the same register as speech fights the dialogue no matter how you mix it.
Energy curves are underused. Phrases like "begins with only bass and a shaker, adds piano after eight bars, full arrangement by the halfway point" give you a structure you can cut against. Even if the model interprets the instruction loosely, you get more usable variation than a flat request for a single vibe.
What to leave out of a prompt
Three things reliably ruin AI music for video:
- Named living artists. Beyond rights concerns, imitation prompts tend to produce generic approximations while inviting legal questions you do not want.
- Stacked, contradictory genres. "Lo-fi orchestral trap jazz with country vocals" averages into a muddy middle.
- Lyrics when you have narration. Sung words and spoken words compete for the same attention band. If you need a vocal texture, ask for wordless female oohs or a hummed melody and keep it low in the mix.
Generate four variations rather than one, listen to them against the picture, and keep the stems if your tool exports them. A separated drum, bass, and melodic layer turns one track into three mixing options.
Voiceover: Casting, Delivery, and Emotional Range
Synthetic narration is now good enough for explainers, documentaries, corporate content, and most social video. It is not equally good at everything, so match ambition to material.
Match the voice to the story, not the trend
Audition voices with your actual script, never with the demo sentence the tool provides. A voice that sounds authoritative reading a product blurb may sound cold reading a first-person story. Consider five attributes: age range, texture (bright, breathy, gravelly), pace, accent, and warmth. Write down the winner's attributes in a voice sheet so a series stays consistent across episodes.
One practical rule: for narration over busy visuals, choose a slightly slower voice with lower pitch variance. For high-energy short-form content, choose a voice that already sounds animated so you do not have to push the speed setting.
Direct delivery with punctuation, pauses, and markup
Punctuation is direction. Commas create micro-pauses, periods create full stops, em dashes create interruption, and paragraph breaks create breath. If you write one long sentence with no internal punctuation, you will get one long unpunctuated delivery.
Most tools also accept some form of markup or parameter control:
- Rate: keep narration between 140 and 160 words per minute for comprehension, faster only for list-style content.
- Pitch and emphasis: use sparingly; heavy emphasis sounds like a commercial read.
- Breaks: insert explicit pause tags of 200-400 ms at structural transitions.
- Pronunciation: maintain a small dictionary for brand names, acronyms, numbers, and place names. Test each one once, then never think about it again.
Numbers deserve special attention. Decide once whether "1,200" is read as "one thousand two hundred" or "twelve hundred," then write it that way in the script.
Multi-speaker scenes and localized versions
For dialogue, generate each speaker separately and cut between them on the timeline. Generating two voices in one pass almost always blurs their timing and makes editing impossible. Keep each speaker on their own track so you can adjust levels independently and add slight panning for a conversational feel.
For localization, do not translate word-for-word and hope the timing holds. Adapt the script to the target language's natural sentence length, then re-time the read against the picture. Some languages expand by 20 to 30 percent. If on-camera speakers need to match the new audio, lip-sync tools can help, but expect to shorten lines rather than stretch them.
Syncing Generated Audio to Picture
Generation produces a file. Editing produces a soundtrack. The gap between them is sync work.
Hit points, temp tracks, and iterative trimming
Mark your hit points before you touch the music: cuts, reveals, punchlines, and transitions. Then place the musical events you generated so they land on those points. If a drum fill arrives half a second after a cut, nudge the audio, not the video, and only if the picture can tolerate the shift.
The temp track method saves time. Cut the video against temporary library music with the right tempo and energy, lock the edit, then generate a custom track to that tempo map. You already know what the music needs to do because you have heard something do it.
Extending, looping, and stitching short clips
Many generators return clips of 30 to 60 seconds. For longer pieces, use whichever of these fits the material:
- Extend or continue features, which preserve instrumentation and key across a longer render.
- Looping a stable section, typically eight or sixteen bars, with the loop point placed on a downbeat.
- Stitching two beds at a structural transition, with a 20 to 40 ms crossfade at a zero crossing to avoid a click.
- Layering: a rhythmic bed under a sustained pad, faded in at the moment the story shifts.
Always leave a clean tail. A track that ends mid-phrase undermines an otherwise tidy ending.
Mixing and Loudness: Making It Sound Intentional
Mixing AI audio is mostly about relationships: voice to music, music to ambience, and the whole mix to the delivery platform.
Rough loudness targets by destination
- Social and web video: roughly -14 LUFS integrated, true peak no higher than -1 dBTP.
- Podcast and spoken-word audio: around -16 LUFS integrated.
- Broadcast-style delivery: around -23 LUFS integrated with tighter true-peak limits.
- Cinema: wider dynamic range with dialogue anchored well below the loudest effects.
These are starting points, not laws. What matters is consistency across a series and headroom before the limiter.
Ducking, EQ carving, and dialogue clarity
Sidechain compression is the workhorse. Route the music to a bus, trigger a compressor on that bus from the voice track, and aim for 3 to 6 dB of reduction with a fast attack and a 150 to 300 ms release. The music should feel like it steps back, not like it is being pumped.
Ducking alone is not enough. Carve a gentle 2 to 3 dB dip in the music between roughly 1 kHz and 4 kHz, where speech intelligibility lives. If your generator exports stems, simply lower the melodic layer and keep the percussion and bass — arrangement solves problems that EQ cannot.
Stereo width and mono compatibility
A surprising share of viewing happens on a phone speaker that is effectively mono. Check your mix in mono: if the voice thins out or the music collapses, you have phase problems, usually from wide stereo reverb on the narration. Keep low frequencies centered below about 120 Hz, narrow the voice, and reserve wide stereo movement for the music bed.
Also check ambience continuity. Absolute silence between lines sounds like a broken file. A quiet room tone at -45 to -50 dBFS holds the scene together and makes edits invisible.
A Quality-Control Checklist Before You Publish
Run the same pass on every project. It takes ten minutes and catches nearly everything.
- Sync: check three points — first line, midpoint, final line — against lip movement or visual cues.
- Intelligibility: listen once on phone speakers, once on earbuds, once on something with bass.
- Loudness: confirm integrated level and true peak for the destination platform.
- Mono check: fold the mix to mono and verify the voice survives.
- Music under dialogue: confirm no melodic line competes with speech at any point.
- Pronunciation: verify names, numbers, acronyms, and non-native words.
- Captions: confirm caption timing matches the audio, not the original script.
- Edges: trim silence at the head, keep a short tail, and avoid starting mid-word.
- Ambience: confirm no dead silence between lines.
- Assets: export the music bed, voice tracks, and mix separately for future revisions.
Common Mistakes That Wreck AI Sound
Generating music before locking the edit. You end up cutting around the track instead of the story. Lock picture first.
Asking for one long perfect track. Structure beats perfection. Generate sections and blend them.
Narrating too fast. Anything above roughly 170 words per minute loses comprehension on mobile, where attention is already thin.
Ignoring sibilance and plosives. Synthetic voices often over-articulate S sounds and P sounds. A de-esser and a gentle high-pass solve most of it.
Layering sung vocals under narration. Two competing word streams make both harder to follow. Use instrumental beds or wordless textures.
Over-compressing to fake loudness. Heavy limiting flattens the dynamics that make a soundtrack feel cinematic and makes narration fatiguing.
Forgetting mono. Half of your audience is listening on a device that cannot reproduce your stereo field.
Reusing one voice and one track across an entire series. Consistency is good; monotony is not. Rotate two or three beds and vary pacing between episodes.
Skipping the script read-through. Most bad narration is a writing problem that no voice setting can fix. Read the script aloud before generating a single line.
Rights, Disclosure, and Working Responsibly
Generated audio raises practical questions that are easier to handle before publication than after.
Terms of use. Commercial permissions differ between tools and can change. Read the current terms for each generator you rely on, and note whether commercial use, monetization, or redistribution is permitted for the outputs you plan to publish.
Voice cloning consent. Only clone a voice you own or have explicit written permission to clone. This applies to your own voice used on someone else's behalf as well.
Style imitation. Prompts that name a living artist or ask for a recognizable voice are a bad idea on both ethical and practical grounds.
Disclosure. Platform rules and audience expectations vary. Documentary, journalism, and educational contexts generally benefit from a clear statement that narration or music is synthetic; entertainment content often does not require it. When in doubt, a short on-screen note or a line in the description removes ambiguity.
Record keeping. Maintain a simple manifest for each project: asset name, tool, model version, prompt, and generation date. If a question arises months later, you can answer it in minutes instead of reconstructing it from memory.
FAQ
Can I publish AI-generated music in monetized videos?
Usually yes, if the tool's terms grant commercial rights to outputs and you are not reproducing identifiable copyrighted material. Verify the specific terms, keep your generation records, and avoid imitation prompts. If a platform's content policy requires disclosure of synthetic media, follow that policy.
Is synthetic narration good enough for professional work?
For explainers, tutorials, corporate video, and most social content, yes — provided the script is written for the ear and the voice is cast deliberately. For emotionally complex acting or anything where vocal nuance carries the message, a hybrid approach works better: use synthetic narration for the bulk and record a human for hero lines.
How long should a generated music bed be?
Match the section, not the video. Beds of 15 to 40 seconds that map to structural beats are easier to mix than a single four-minute track. Stitch them at transitions and crossfade briefly.
How do I keep music from drowning out narration?
Combine three techniques: choose an arrangement with space in the speech range, duck the music 3 to 6 dB under the voice with a fast-attack sidechain, and carve a shallow dip between roughly 1 kHz and 4 kHz on the music bus.
Should I always disclose that the audio is AI-generated?
Not always, but err toward transparency when the audience might reasonably assume a human speaker or performer — journalism, testimonials, education, and anything presented as documentary evidence. Check the platform's synthetic-media policy as well.
What about languages and accents?
Quality varies widely by language, region, and voice model. Test every name and number in the target language, prefer native-quality voices for hero content, and keep a pronunciation list. For heavily localized projects, consider a native speaker for final review, since a slightly wrong stress pattern is more noticeable than a slightly wrong accent.
How many music variations should I generate?
Four is a reasonable default: two that follow the brief closely and two that push the tempo or instrumentation in a different direction. Audition them against the actual cut rather than in isolation, and keep one spare for revisions.
