Audio Is Half the Edit, and Usually the Half That Gets Rushed
Viewers will forgive soft focus, a slightly crooked horizon, or a jump cut that lands a beat late. They will not forgive audio that is muddy, uneven, or impossible to follow. Audio carries comprehension, pacing, and emotion simultaneously. It tells the audience when to laugh, when to lean in, and when a scene has ended. When the sound is wrong, everything above it in the frame looks wrong too.
For years, decent audio meant either a booth, a microphone, and a patient narrator, or a licensing budget for stock music. That bottleneck has largely collapsed. Synthetic narration and generated music are now good enough for real production work: explainers, product demos, course modules, short-form social edits, training videos, and documentary-style pieces. The catch is that generation is the easy part. Arrangement, leveling, consistency, and taste are where most projects still fall apart.
This guide is a practical workflow, not a tour of features. It covers the three audio layers every video needs, how to audition and direct a synthetic voice, how to generate music that actually fits an edit, how to assemble a mix that survives phone speakers, and how to run quality control before you export. The tools change constantly; the workflow stays useful.
The Three Audio Layers in Any Video
Most videos are built from three distinct audio layers. Treating them as one blob is the single most common cause of amateur-sounding results.
Voiceover: the narrative spine
The voiceover carries meaning. It should be intelligible in one pass, with no rewinding required. That means prioritising clarity over character: consistent pace, clean consonants, and room to breathe between ideas. A distinctive voice is a bonus, but an unclear voice is a failure, no matter how stylish it sounds.
Music: the emotional current
Music tells the audience how to feel about what they are seeing. It is not decoration. A gentle piano bed under a product walkthrough signals calm competence; the same footage under a driving synth loop reads as urgent and futuristic. Music also masks edit seams and smooths abrupt visual transitions, which is why a good bed makes an edit feel more polished than it visually is.
Sound design and ambience: the texture
Ambience and effects make a scene feel like a place rather than a graphic. Room tone under an interview, a subtle whoosh on a transition, a keyboard click under a screen recording, a distant crowd under a street shot. These elements are quiet and mostly unnoticed, which is exactly the point. When they are missing, footage feels uncanny; when they are too loud, they feel cheap.
Build them in order: voice first, music second, texture third. Each layer constrains the next. A dense music bed forces you to slow the narration; a sparse ambience gives the voice room to be intimate.
Casting an AI Voice: Criteria Beyond "Does It Sound Real?"
Realism is the lowest bar. Modern text-to-speech clears it easily. What distinguishes a usable voice from an unusable one is how it behaves under pressure: long sentences, numbers, acronyms, emotional shifts, and repeated takes.
Tone, age, and register
Start with the role the voice plays. A narrator who explains software should sound like a knowledgeable colleague, not a movie trailer. A character in a children's story needs warmth and elasticity. A corporate training module needs neutrality, because a strongly characterised voice becomes irritating across forty minutes of material.
Write down three adjectives for the voice before you audition anything: for example, "warm, unhurried, credible." Then audition with those filters active. Without them, every option sounds acceptable and you waste an afternoon.
Pace, pauses, and breath
Generated narration often defaults to a pace that is slightly too fast and pauses that are slightly too short. Test each candidate at your intended playback speed and at 0.75x. If the voice degrades into a robotic slur when slowed, it will degrade in every slow-motion montage too. Listen for breath placement. Natural speakers breathe at clause boundaries; synthetic voices sometimes insert breaths at arbitrary points, which reads as uncanny even to viewers who cannot name the problem.
Accent and localization
If you are producing for multiple markets, decide early whether you are dubbing or re-recording. Re-recording with a native voice for each language usually beats translating a script and forcing one voice model to handle it. Localisation also changes rhythm: a script written for brisk English delivery may need longer sentences in a language with more syllables per idea.
A five-minute audition protocol
Use the same short script for every candidate:
- One plain declarative sentence.
- One sentence containing a number, a percentage, and an acronym.
- One question with rising intonation.
- One emotionally warm line, such as a closing thank-you.
- Your actual opening line, word for word.
Score each candidate on clarity, pace, naturalness, and how much editing you would need to do. The winner is rarely the one that sounds most impressive in isolation; it is the one that requires the least repair.
Writing a Script That a Synthetic Voice Can Perform
A voice model is a performer reading what you wrote. Give it a difficult script and you will spend your time editing rather than creating.
Keep sentences short. One idea per sentence, roughly twelve to twenty words. Long subordinate clauses invite odd intonation, because the model has to guess where the emphasis belongs.
Expand anything ambiguous. Write "twenty-five percent" rather than a percent sign if the model misreads it, and spell out initialisms phonetically the first time if they are unusual. Test every number in your script before you render the full pass.
Use punctuation as direction. Commas create small lifts, periods create full stops, em dashes create a beat of suspense, and ellipses create hesitation. Line breaks between paragraphs create a longer pause than a period. Most voices respond to this more reliably than they respond to emotion sliders.
Read it out loud yourself. If you stumble, the model will too. If you run out of air, so will the listener's attention.
Do the timing math. Narration lands between roughly 130 and 160 words per minute depending on pace and language. A 900-word script at 145 words per minute is about six minutes of narration, which tells you whether your planned runtime is realistic before you generate anything.
Generating Background Music That Matches the Edit
Music generation is where enthusiasm most often outruns judgement. The goal is not the best-sounding track; it is the track that fits this edit.
Start with a temp track, not a prompt
Before generating anything, place any existing track you like under the rough cut and watch it through. Note exactly where the emotion changes: second twelve, second forty, the reveal at one minute. You now have a shot list for your music. Then describe those requirements when you generate, rather than asking for "upbeat corporate music."
Emotional consistency across scenes
A common failure is a generated score that shifts mood every eight seconds because the prompt implied variety. For a single scene, ask for one sustained mood. For a multi-scene piece, generate separate pieces per scene and blend them with crossfades, or ask for one theme with variations in instrumentation and intensity.
Describe your requirements with concrete parameters: instrumentation, tempo range, energy level, whether there is a percussive element, whether there are vocals, and how the piece should end. A useful prompt reads like a brief, not a vibe.
Structure: loops, stems, and levels
Ask for stems when the tool supports them. Separate drums, bass, melody, and pads let you strip the track down under narration and bring it back up in the gaps. This single technique makes generated music sound professionally placed, because the arrangement is responding to the voice instead of competing with it.
Loop points matter for anything longer than a minute. Check that the loop is seamless before you rely on it, or edit the tail manually so the last bar resolves.
Ducking and bed levels
Once the music is in the timeline, set the bed 15 to 22 dB below the voice, then use sidechain or manual automation to duck the music another 3 to 6 dB whenever narration is present. The music should be clearly audible in the gaps and clearly subordinate underneath speech. If you need to strain to hear the narration on a phone speaker, the bed is too loud.
The Assembly Pipeline: From Timeline to Finished Mix
A repeatable order of operations saves more time than any single tool.
- Lock picture first. Never score an edit that is still changing. Moving a shot after you set beats means redoing the music placement.
- Lay a temp score. Use any track that carries the right emotion, purely to establish pacing.
- Generate the narration in segments. Render scene by scene rather than the whole script at once. This limits re-renders, keeps file sizes manageable, and lets you adjust performance progressively.
- Edit the voice like dialogue. Trim leading silence, remove unnatural gaps, and use short crossfades rather than hard cuts between segments. Consistency of room tone between segments matters more than any single line.
- Replace temp music with generated stems. Align the emotional peaks with the visual beats you identified earlier.
- Add ambience and effects on a separate bus. Keep them at least 20 dB below the voice by default; raise only where the scene needs presence.
- Automate ducking. Ride the music down under speech and up in the gaps. Do it by hand for short pieces, with sidechain compression for long ones.
- Check loudness and export. Target roughly -14 LUFS integrated for general web video and around -16 LUFS for podcast-style delivery, with true peaks no higher than -1 dB.
The mono check
Before you finish, fold the mix to mono and listen. A large share of viewers watch on a phone speaker or with one earbud in. If your ambience or a stereo-panned instrument disappears in mono, or if the voice becomes thin, fix it before export.
Quality Control Checklist Before You Export
Run this list every time, even when you are in a hurry.
- Narration is intelligible at low volume on a phone speaker.
- No clipped or distorted consonants, especially on plosives.
- Music never masks a key word; check the busiest ten seconds of speech.
- Ambience is present but not identifiable as a loop.
- Every generated segment matches the neighbours in tone and room character.
- Numbers, names, and brand terms are pronounced correctly throughout.
- Loudness and true peak targets are met.
- A mono fold-down still sounds balanced.
- Music and voice are exported as separate stems for future re-edits.
The last point is easy to skip and expensive to regret. Separate stems let you re-version a video for a new market or platform without regenerating everything.
Common Mistakes and How to Fix Them
Generating the full script in one pass. One mispronounced word forces a full re-render, and you lose the ability to tune performance scene by scene. Fix: render in paragraph or scene chunks.
Choosing a voice in isolation. A voice that sounds great reading a sample line can be exhausting across a long explainer. Fix: audition with your own script and your own runtime.
Letting music set the mood alone. If the music is doing all the emotional work, the visuals and script are probably vague. Fix: write the intended emotion into the script first, then support it with music.
Ignoring the first two seconds. Many platforms autoplay muted, so the opening must work visually, but the first audible moment still decides whether viewers keep sound on. Fix: front-load a clean, confident line rather than a long musical intro.
Mixing on headphones only. Headphones flatter everything. Fix: always check on a phone speaker and one earbud.
No naming convention. Files like final_v3_final.wav create chaos across versions. Fix: name by project, scene, layer, and version, and keep a single audio folder structure.
How to Choose Tools Without Getting Locked In
Feature lists are less useful than five practical questions.
Language coverage. Does the voice set support every market you actually serve, with native-sounding pronunciation rather than a heavy accent in a second language?
Export quality. Can you get uncompressed 48 kHz WAV, and separate stems for music? Lossy exports will haunt you in the final mix.
Re-render behaviour. Can you regenerate one segment without changing the rest? Deterministic, segment-level control is worth more than a larger voice catalogue.
Rights clarity. Confirm that commercial use, monetisation, and client work are covered by the terms you agreed to, and keep your own record of which asset came from which tool.
Automation. Batch generation, script files, and API access matter if you produce more than a couple of videos a month. Manual interfaces become the bottleneck fast.
A sensible stack is often two tools: one focused on narration with strong pacing control, and one focused on music with stem export. Keeping them separate makes replacement easy when one improves or changes its terms.
FAQ
Can synthetic narration pass as human narration?
For informational and instructional content, yes, most of the time. Listeners notice unusual breathing, flat emotional peaks, and unnatural number reading far more than they notice the synthesis itself. Fixing those three things improves perceived humanity more than switching models.
How do I stop music from burying the narration?
Set the bed 15 to 22 dB below the voice, then duck it a further 3 to 6 dB under speech. Use stems so you can thin the arrangement during dense passages rather than lowering the whole track, which makes the music sound distant and lifeless.
What loudness should I target?
Around -14 LUFS integrated for general video platforms and about -16 LUFS for spoken-word podcast delivery, with true peaks at or below -1 dB. Platforms normalise anyway, so consistency across your own catalogue matters more than hitting an exact number.
Do I need separate tools for voice and music?
Not strictly, but specialisation usually wins. Narration tools optimise for pacing, pronunciation, and re-rendering; music tools optimise for stems, structure, and tempo control. If a single tool does both well and exports clean stems, that is a legitimate reason to consolidate.
How do I keep audio consistent across an episodic series?
Freeze a template: same voice, same pace setting, same music instrumentation family, same ambience set, same loudness target, same export settings. Consistency across episodes is a branding decision, not an artistic one, and templates make it automatic.
How much time should a five-minute video's audio take?
With a locked picture and a prepared script, roughly one to two hours: thirty minutes for narration generation and editing, twenty for music placement, twenty for ambience and ducking, and twenty for the quality control pass. Scrambling for a voice or fixing a bad script mid-edit is what turns that into a full day.
