Viewers forgive soft focus, shaky handheld shots, and slightly off white balance. They do not forgive dialogue they cannot understand. Audio is the quiet quality signal in video production: when it is clean, nobody mentions it, and when it is muddy, no amount of colour grading or clever editing rescues the piece.
This guide walks through a practical voice studio workflow for AI narration and background music. It covers script preparation, voice selection, prosody control, generative music, ducking, loudness targets, and the checks worth running before you publish. Everything here is tool agnostic enough to apply whether you edit in a full NLE or a browser-based timeline.
Why Audio Decides Whether a Video Feels Professional
Listeners process speech and music through the same limited channel. When narration competes with a music bed, the brain spends effort separating the two streams instead of absorbing the message. That effort shows up as drop-off in the first thirty seconds, and it is far more common than creators expect.
Three layers carry almost every video: voice, music, and effects or ambience. Each has a job. The voice carries meaning and must stay intelligible at any playback volume, from phone speakers to noise-cancelling headphones. Music carries emotion and pacing, signalling when a section is tense, warm, or triumphant. Effects and ambience carry place, telling the viewer whether a scene happens in a kitchen, a server room, or a windy street.
Problems appear when the layers swap jobs. Music that tries to carry meaning drowns the voice. Effects that try to carry emotion become distracting. A voice that tries to carry energy with constant volume becomes exhausting. Deciding early which layer owns which responsibility is the single most useful planning step in an audio workflow.
The second reason audio matters is consistency. A channel with ten videos in slightly different acoustic spaces feels amateurish even if each video is fine on its own. Standardising voice, loudness, and music style creates a recognisable sound identity, and that identity compounds the way a visual style does.
The Modern Audio Stack: What Each Layer Does
A modern production stack has four functional layers. Understanding them separately makes troubleshooting much faster, because you can isolate which layer introduced a problem instead of tweaking everything at once.
Script and pronunciation preparation
This layer is text work, not audio work, but it determines roughly half of the final quality. Scripts written for the eye and scripts written for the ear are different documents. Short clauses, repeated key nouns, and explicit transitions all help synthetic voices sound natural because the model receives cleaner structure to render.
Preparation also means deciding pronunciation before generation: how acronyms are spoken, whether numbers are read as digits or words, and which brand names need special handling. Building a small pronunciation sheet for recurring terms saves hours across a series.
Text-to-speech and voice cloning
Modern TTS engines generate speech with controllable timbre, pace, and emotional colour. The practical controls that matter most are stability or expressiveness sliders, pace, pitch offset, and pause insertion. Consistency controls, sometimes called seeds or voice presets, matter more than raw realism for episodic content because they keep the same character recognisable across recordings.
Voice cloning deserves a short caution. Only clone voices you own or have documented permission to use, and keep the consent record with the project files. Synthetic narration can also require disclosure depending on platform policy and local regulation, so check before publishing anything that could be mistaken for a real person speaking.
Generative music and ambience
Text-to-music models have become genuinely useful for beds, transitions, and stingers. You describe mood, instrumentation, tempo, and energy, and the model returns a track or a set of stems. The limitation is structure: generated tracks rarely follow your edit, so plan to cut, loop, and re-arrange rather than drop in a full track from start to finish.
For ambience, generative tools are excellent. Room tone, crowd murmur, rain, and machinery hum are the kinds of textures that make a scene feel located, and they are tedious to source by hand.
Mixing, ducking, and mastering
The final layer sets relationships. Level balance, frequency carving, dynamic ducking, and loudness normalisation all live here. This is where the voice stops being a file and becomes a soundtrack. Most creators underinvest in this layer, and it is the fastest place to gain perceived quality.
Choosing the Right Voice: A Practical Decision Framework
Auditioning voices by vibes alone leads to rewrites. Use a short set of criteria instead, and score three candidates against them.
Timbre and register
Low, warm registers read as authoritative and calm. Mid-range voices read as friendly and conversational. Bright, higher registers read as energetic and youthful. Match the register to the emotional default of your content, not to your own voice. A meditation channel with a bright, fast voice will fight its own subject matter.
Also check the frequency range of the voice against your music. Voices with heavy low-end content clash with bass-heavy beds. Voices with a lot of 2 to 5 kHz energy cut through almost any bed, which is convenient for documentary and explainer work.
Pace, pauses, and prosody
Generate the same two sentences at three pace settings and listen for three things: whether pitch stays level at the end of statements, whether commas produce real breaths rather than clipped gaps, and whether the voice stresses the words you would stress. Prosody is where synthetic narration usually reveals itself, and it is also the easiest thing to fix with punctuation and explicit pause tags.
If a model supports emotional intensity controls, test the middle of the range first. Maximum expressiveness often produces theatrical results that feel wrong in instructional content, while minimum expressiveness produces flat, robotic delivery.
Accent, language, and localisation
If you publish in more than one language, decide early whether to localise with native-language voices or a single voice speaking multiple languages. Native voices usually sound better and convert better, but they cost more production passes. A hybrid approach works well: one primary voice per language, with a shared music and loudness template so the channel still feels unified.
Check accent fit too. A generically neutral accent is safest for broad audiences, while a strong regional accent can build intimacy with a specific audience. Neither is inherently better; the mistake is choosing one by accident.
Writing Scripts That Synthesize Well
Synthetic narration amplifies whatever you give it. Sloppy punctuation becomes strange pauses, and ambiguous abbreviations become confident mispronunciations.
Punctuation as performance direction
Treat commas as short pauses and full stops as longer resets. Break long sentences into two or three shorter ones. Where a model supports pause tags or break durations, use them explicitly for dramatic beats rather than relying on ellipses.
Avoid excessive parentheticals, em dashes that confuse chunking, and long lists without separators. If a sentence requires a breath in the middle, add a comma and regenerate.
Normalising numbers, units, and acronyms
Decide a house style and apply it consistently. Write out small numbers as words when they should be spoken naturally, keep digits where they should be read as figures, and spell acronyms phonetically when the model guesses wrong. Units should be expanded, so users hear kilometres per hour rather than a jumble of letters.
Build a shared glossary for recurring terms: product names, place names, job titles, and technical vocabulary. Paste it into your notes app and reuse it for every episode.
Pronunciation dictionaries and retakes
Most serious TTS tools allow phonetic overrides for individual words. Use them instead of rewriting whole sentences. When only one word fails, fix that word; when a whole clause fails, regenerate the clause. Chunking scripts into paragraphs and generating them separately gives you surgical retakes without touching the rest of the take.
Keep the raw generated files. A clean retake months later still needs to match the original session, and regenerating from scratch rarely matches exactly.
Background Music: Generation, Selection, and Structure
Music is the fastest way to change how a video feels and the easiest way to ruin clarity. Think of it as a pacing instrument rather than wallpaper.
Matching genre to intent
Start from the emotional verb of the section. Is it explaining, reassuring, celebrating, or warning? Each maps to a different palette. Sparse piano and warm pads suit explanation. Light percussion and plucked strings suit momentum. Low drones suit tension. Aggressive drums suit only genuine climaxes.
Tempo matters as much as genre. Music above roughly 120 BPM pushes the edit forward and can make calm narration feel rushed, while music under 80 BPM can make energetic sequences drag.
Building an arc with stems and loops
Request stems when the tool offers them: drums, bass, harmony, melody. With stems you can enter a section with only harmony, add percussion when the topic intensifies, and drop everything for the closing sentence. That single technique makes generated music feel composed rather than pasted.
Loop points need attention. Find a bar boundary where the energy is stable and cut there, then crossfade by 20 to 50 milliseconds to hide the seam. Avoid looping sections that contain a rising fill or a vocal chop, because both announce the repeat.
Ambience and effects as glue
A quiet ambience bed under the music makes edits feel continuous. Room tone fills the silence between sentences so cuts do not create dead air. Two or three subtle transition sounds, used consistently, give the video a signature without becoming gimmicks.
Keep effects short and low. If a swoosh draws attention to itself, it is too loud.
The Mix: Levels, Ducking, and Loudness Targets
Good mixing is mostly subtraction. You are removing competition so the voice can win without being loud.
Ducking that stays musical
Ducking lowers music automatically whenever narration plays. Sidechain compression is the classic approach: feed the voice channel into the compressor on the music bus. Reasonable starting points are a 4:1 ratio, 6 to 9 dB of gain reduction, a 5 to 10 millisecond attack, and a 250 to 400 millisecond release. A release that is too fast makes music pump audibly; one that is too slow makes the voice sound buried at the start of each sentence.
For narrated explainers, gentle ducking plus manual volume automation beats aggressive ducking alone. Write keyframes manually at section boundaries so the music breathes in gaps instead of staying permanently quiet.
EQ carving instead of volume wars
The voice lives mostly between 100 Hz and 8 kHz. Music beds usually carry energy in the same range. Rather than turning music down until it disappears, carve a shallow dip of 2 to 3 dB in the music between roughly 1 and 4 kHz where the voice has presence. High-pass the voice at 80 to 100 Hz to remove rumble and plosive energy. On the music, consider a high-pass around 40 Hz unless you deliberately want sub-bass weight.
Control sibilance with a de-esser rather than a broad high-frequency cut, which dulls the whole voice. Reserve narrow cuts for specific problems and keep processing light; stacked plugins add artefacts faster than they add clarity.
Loudness targets by platform
Integrated loudness targets differ across platforms and distribution. Common reference points are around -14 LUFS for video streaming, -16 LUFS for podcasts and spoken-word audio, and -23 LUFS or -24 LKFS for broadcast standards. Set true peak ceilings at about -1 dBTP to leave headroom for lossy encoding.
Rather than mastering by ear for each destination, export one dynamic master and generate normalised deliverables from it. Consistency across your catalogue matters more than hitting an exact number, because platforms will adjust anyway.
A Repeatable End-to-End Workflow
Lock the script. Proofread for ear-friendly phrasing and pronunciation before generating anything.
Audition three voices. Score them on register, pace, and clarity against a music bed, not in isolation.
Generate in paragraphs. Keep chunks small enough for surgical retakes and consistent enough for a stable character voice.
Clean the voice track. Remove breaths that break rhythm, tame sibilance, and apply gentle noise reduction only where needed.
Map the music. Sketch a mood and energy curve across the timeline before choosing tracks or prompts.
Assemble with stems. Bring music elements in and out at section boundaries rather than playing one track end to end.
Set levels and ducking. Balance voice, music, and ambience, then add sidechain control and manual automation.
Master to target. Apply gentle bus compression, loudness normalisation, and a true peak ceiling.
Listen on three systems. Phone speaker, laptop speaker, and headphones. Fix anything that fails on the smallest one.
Archive stems and settings. Save the raw voice takes, music stems, and bus presets so the next episode starts from a known good state.
Quality Control Checklist Before You Publish
Run this list on every export, in order. It takes five minutes and catches most embarrassments.
- Every word of narration is intelligible on a phone speaker at 50 percent volume.
- No sentence is clipped at the start or end.
- Music never masks the first words of a new section.
- There are no abrupt silence gaps where edits meet.
- Loudness is within about 1 LU of your catalogue standard.
- True peak stays below the ceiling after encoding.
- Mouth noise, clicks, and breath spikes are controlled but not sterilised.
- Ambience continues under transitions instead of stopping dead.
- Captions or subtitles match the spoken audio, including numbers and names.
- The file name and version number follow your naming convention.
Common Mistakes and How to Avoid Them
Over-processing the voice. Heavy compression, aggressive de-essing, and stacked EQ moves create a thin, metallic sound. Make one change at a time and compare against the raw take.
Mixing only on headphones. Headphones hide how badly a mix collapses on small speakers. Always check the smallest playback device you expect people to use.
Letting music set the pace for the whole video. A single track for eight minutes becomes wallpaper. Vary density, not just volume.
Inconsistent voices across a series. Episode two should sound like episode one. Save presets, seeds, and pace settings in a project document.
Ignoring loudness normalisation. Platforms turn quiet uploads up and loud uploads down, so extreme levels only hurt dynamics. Master to a sane target and stop worrying about perceived volume wars.
Skipping the silent listen. Play the timeline with the music muted. If the narration alone is boring or unclear, music will not fix it.
Forgetting accessibility. Captions, transcripts, and clear speech help more of your audience than any stylistic choice. Treat them as part of the audio workflow, not an afterthought.
FAQ
How long does a typical voice and music pass take? For a five-minute explainer, expect one to two hours once your presets exist: around thirty minutes for script preparation and generation, twenty for voice cleanup, thirty for music assembly, and twenty for mixing and mastering. The first project in a new format takes considerably longer because you are still defining the template.
Should I generate one long voice file or many short ones? Many short ones. Paragraph-level chunks give you cleaner retakes, let you fix single mispronunciations, and make it easy to reorder sections later. Assemble them on the timeline with small overlaps rather than gluing them into a single file first.
Is generated music safe to publish? It depends on the terms of the tool you use. Check whether commercial use is permitted, whether attribution is required, and whether your inputs or outputs may be used for training. Save a copy of the terms and the generation date alongside the project files.
How loud should background music be under narration? As a rough guide, music should sit well below the voice and rise noticeably only during gaps, intros, and outros. If you can comfortably follow the narration without effort, the balance is probably right. If you have to concentrate, pull the music down 2 to 3 dB and re-check on a phone speaker.
Do I need a separate mix for social clips? Yes, usually. Vertical short-form is watched on phone speakers in noisy environments, so dialogue clarity matters more and dynamic range matters less. Create a slightly denser, more compressed version with less ambience, and keep the same loudness target for consistency.
What if my synthetic voice sounds robotic no matter what I change? Change the source text before changing settings. Break long sentences, add explicit pauses, stress key words by repeating them, and remove nested clauses. If it still fails, switch voices; some registers simply do not suit certain subject matter.
Bringing the Workflow Together
A reliable audio pipeline is less about finding the perfect voice model and more about building repeatable decisions: which layer owns which job, how loud each layer sits, and what checks run before export. Once those decisions are written down, each new video becomes an assembly job rather than an experiment.
Start small. Pick one voice, one music palette, and one loudness target. Generate a two-minute test, mix it, and listen on a phone. Then refine the template rather than rebuilding it, and let consistency do the work that individual tweaks cannot.
