Why Audio Decides Whether a Video Feels Professional
Viewers forgive a lot. They forgive soft focus, a slightly shaky handheld shot, a colour grade leaning warm. They rarely forgive bad audio. When dialogue sits under a louder music bed, when a synthetic narrator mispronounces every third word, or when room tone vanishes between cuts, people leave — usually without being able to explain why. Audio is the invisible half of video, and it is the half that decides whether a project feels amateur or authored.
The reason is asymmetry. The brain parses visuals quickly and redundantly; a few dropped frames go unnoticed. Speech is processed in real time with far less tolerance. If a listener has to work to separate words from background noise, that effort competes with comprehension itself. A clean voice track over modest footage holds attention longer than a cinematic montage with muddy sound.
It also compounds over time. A series with consistent audio builds trust; viewers stop bracing for problems and start following the argument. That is why producers who work with generated voice and generated music obsess over the unglamorous parts — levels, pauses, transitions — rather than the spectacular ones.
Treat audio as a production stage rather than a final polish step. Decide how the piece should sound before deciding how it should look.
The Modern Audio Stack, Piece by Piece
Generative audio is not one technology; it is a small stack of specialised systems. Knowing which layer you are working in makes troubleshooting much faster.
Voice synthesis
Text-to-speech pipelines convert text into phonemes, map them to an acoustic representation, and render that into a waveform with a vocoder. Contemporary models are trained on many hours of speech and capture breath, micro-pauses, and subtle pitch drift. Most interfaces expose pace, pitch, energy, and sometimes explicit emotion tags. Some accept a short reference clip to imitate a particular voice.
The catch is consistency. A voice that sounds warm in one paragraph can drift in the next if settings change. Fix your parameters early and record them somewhere permanent.
Music generation
Text-to-music models are conditioned on descriptions of genre, instrumentation, tempo, and mood. Outputs arrive as stereo mixes or, better, stem groups — drums, bass, harmony, pads, melody. Stem export is the feature that matters most, because it lets you rebalance a track after the fact instead of regenerating it from scratch and hoping.
Mixing and mastering utilities
The finishing layer handles repair and loudness: noise reduction, de-reverb, plosive repair, stem separation, and loudness matching across clips. These are unglamorous but they determine whether your edit sounds like one coherent piece or five unrelated files stitched together.
How the layers interact
Each layer introduces artefacts the next one must clean up. Aggressive music compression makes ducking harder later. Heavy de-noising on the voice removes breath and leaves a synthetic sheen. Render and latency also matter: if a single line takes minutes to produce, your editing rhythm changes, and you start accepting takes you would normally reject. Build a workflow that keeps retries cheap.
Scripting and Directing Synthetic Voice
Punctuation is direction
Synthetic voices read punctuation literally. A comma is a short pause, a period a longer one, an em dash a break in thought, an ellipsis hesitation. If you want a pause that the grammar does not justify, insert a line break or a spaced dash instead of relying on intuition.
Spell for sound
Write numbers, acronyms, units, and links the way you want them spoken. A four-digit number may be read digit by digit in one context and as a quantity in another. Abbreviations get expanded unpredictably. Homographs — read, lead, live, wind, bass — are a reliable source of embarrassing errors. Keep a pronunciation note per project and update it every time something goes wrong.
Keep lines short and breathable
Long sentences force flat delivery. Break narration into lines of twelve to eighteen words, one idea each. When a line is short, a model naturally places emphasis; when it is long, everything gets equal weight and the result sounds like a machine reading a manual aloud.
Build a voice profile
Record model name, voice identifier, pace, pitch, energy, and gain. Reusing the exact profile across episodes is the only dependable way to keep a character recognisable. Save a reference audio file alongside the notes so a future session can be checked against it by ear.
Localisation
Translate meaning rather than syntax, then re-time the result. Idioms, dates, currency formats, and honorifics shift between languages, and a literal translation often produces cadences that no native speaker would use. Test names and brand terms separately; they are where localisation breaks first.
Generating Music That Sits Under a Voice
Describe production, not emotion
Calling a track sad tells a model almost nothing. A description such as solo cello, slow tempo, minor key, no percussion, long reverb tail, low energy, sparse arrangement tells it plenty. Replace adjectives with instrumentation, register, tempo, density, and production era.
Ask for structure and stems
Request sections — intro, build, main, break, outro — rather than a single continuous bed. Ask for stems so you can drop the melody during narration and bring it back on transitions. If stems are unavailable, request an instrumental with no lead melody; that version is almost always easier to duck.
Loop and edit hygiene
Generated tracks often fade oddly at the edges. Trim to a musically sensible point, run a short crossfade, and listen for clicks at the seam. Never let a track restart abruptly mid-sentence, and avoid cutting on a downbeat unless the cut is intentional.
Write music that leaves room
The best background score has a hole where the voice sits. Avoid busy mid-range textures, continuous ride cymbals, and melodic lines in the speech band. If a track sounds slightly thin when played alone, it will usually sound right once narration sits on top of it.
Reference-driven iteration
Describe a track you admire in structural terms — instrumentation, tempo, production era, energy curve — rather than naming it. Generate three variants, then keep the one with the least intrusive melodic content. Melody is what competes with language.
The Mix: Turning Generated Clips Into Coherent Audio
Gain staging first
Set voice peaks around minus six decibels before mastering. Leave headroom rather than pushing everything to the ceiling and relying on a limiter later. Gain staging problems turn into frequency problems once compression is applied.
Carve frequency space
High-pass narration around eighty to one hundred hertz to remove rumble. Reduce the two to four hundred hertz region to clear boxiness. A gentle lift around three kilohertz restores intelligibility, and a narrow cut near seven kilohertz tames sibilance. On the music bus, apply a broad dip between one and four kilohertz — exactly where speech lives.
Duck with intent
Sidechain compression of three to six decibels with a medium release keeps music audible between sentences and out of the way during them. For long narration blocks, automate a static level drop instead; it sounds smoother than pumping compression and is easier to undo.
Match the space
If the visuals show a small room, a long hall reverb on the voice will feel wrong. Keep narration dry or lightly treated and let the music carry the atmosphere. Add subtle room tone beneath edits so silence does not read as a dropout.
Loudness and true peak
Aim for an integrated loudness around minus fourteen LUFS for general web video and closer to minus sixteen LUFS for spoken-word podcasts. Check true peak near minus one decibel to avoid distortion on consumer playback. Normalise clips before editing rather than nudging levels by ear mid-timeline.
Check on the worst speaker you own
A phone speaker exposes low-end buildup and buried consonants instantly. Summing to mono reveals phase problems between duplicated stereo tracks. Run both checks before publishing, not after.
A Step-by-Step Production Workflow
Step 1 — Pre-production map
Write the script in short lines, mark pauses, tag pronunciation risks, and sketch a mood map with timings: where music enters, where it drops out, where a sound effect marks a cut.
Step 2 — Generate and audit the voice
Generate in blocks, then listen at normal speed with your eyes closed. Fix pronunciation first and pacing second. Export takes with descriptive names so you can find them again after a week away from the project.
Step 3 — Music pass
Generate two or three candidates per mood. Drop them under the voice before deciding; tracks that sound impressive alone frequently fail underneath speech.
Step 4 — Detail layer
Add transitions, whooshes, and room tone. Keep this layer quiet. Its job is continuity, not spectacle.
Step 5 — Mix and master
Balance voice, music, and effects, then treat the master bus gently. Compression ratios around two to one and modest limiting preserve the dynamics that make narration feel alive.
Step 6 — Quality control sweep
Listen on headphones, a phone, and a laptop. Check the first fifteen seconds and the final fifteen seconds carefully, since they shape retention. Verify loudness consistency across the whole piece and confirm that captions match the spoken words exactly.
Common Mistakes That Ruin AI Audio
- Music louder than the voice. The most common error by far; if you cannot hear every consonant on a phone speaker, remix.
- No room tone. Hard silence between clips reads as a technical fault rather than a dramatic pause.
- One long take with no variation. Pace variation is what makes narration feel human.
- Ignoring breath. A little breath noise adds credibility; total absence sounds synthetic.
- Mismatched reverb between clips. Different spaces inside one scene break the illusion.
- Regenerating instead of editing. A one-second trim frequently beats a fresh generation.
- Guessing loudness by ear. Measure it and compare across episodes.
- Skipping the mono check. Phase problems hide in stereo and surface on phones.
Rights, Ethics, and Disclosure
Voice cloning requires clear consent from the person whose voice is used, ideally written and time-limited. Impersonating a public figure, even as a joke, can create legal exposure and platform penalties. When a synthetic performer speaks, some platforms and jurisdictions expect disclosure; a brief on-screen note or a line in the description is usually enough.
For music, read the terms attached to the generator you use. Some tools grant broad commercial use of outputs, while others restrict redistributing the audio itself as a standalone product. Keep a simple manifest per project: tool, model, date, prompt, and the licence terms you relied on. If a client asks later, you can answer in seconds instead of reconstructing the details from memory.
Also mind cultural context. A voice that reads as authoritative in one market can read as comedic in another, and music that signals celebration in one region can signal mourning elsewhere.
How to Choose the Right Audio Tools
Judge generators on the things that affect your edit, not the demo reel:
- Control granularity: can you set pace, emotion, and pauses, or only paste text?
- Stem export: can you separate music elements for mixing?
- Consistency: can you reuse a voice or style across sessions with a saved profile?
- Language coverage: does it handle your target languages with native-sounding accents?
- Editing integration: does it export clean WAV files or only compressed previews?
- Batch and automation: can you process many lines without clicking through each one?
- Licensing clarity: are commercial rights explicit and easy to document?
- Speed: how long does a full episode take to render, including retries?
Most teams settle on two voice tools and one music tool — a primary and a fallback for when a phrase simply refuses to render well.
FAQ
How do I stop music from fighting the narration?
Carve a modest dip in the music between roughly one and four kilohertz, keep the music well below the voice in level, and duck it three to six decibels under speech. If it still competes, the arrangement is too dense — regenerate with fewer mid-range instruments.
Can synthetic narration sound genuinely human?
Yes, within limits. Short lines, varied pacing, a little breath, and natural punctuation get you most of the way. Long uninterrupted paragraphs rarely do. If a line matters emotionally, generate several takes and choose by ear rather than by prompt.
Should I generate music before or after the voice?
After. The voice establishes rhythm, sentence lengths, and pauses, and the music must serve that timing. Composing first almost always means re-editing the score later.
How loud should a finished video be?
Around minus fourteen LUFS integrated for general web video, with true peaks near minus one decibel. Spoken-word audio often sits closer to minus sixteen LUFS. The exact number matters less than consistency across a series.
What is the fastest way to fix a mispronounced word?
Regenerate only that sentence, matching the original settings, then splice it in with short crossfades at both ends. Re-rendering the whole script risks changing tone and timing elsewhere.
Do I need to disclose that the audio is AI-generated?
Practices vary by platform and region and are still evolving. Disclosing synthetic voice in the description or on screen is a safe default, especially for anything presented as documentary or news.
What is the biggest quality win for the least effort?
Level matching. Normalising every voice clip to the same target before editing removes most of the perceived unevenness in a long piece, and it takes minutes.
