Audiences forgive a lot of visual imperfection. A slightly soft lens, a slightly warm white balance, a frame that drifts half a degree off level — most viewers never notice. Audio is different. Muffled dialogue, a music bed that fights the narration, or a sudden drop into silence between cuts reads as amateur immediately, even to people who could not explain why.
That asymmetry is why sound design has quietly become the highest-leverage part of small-team video production. A camera upgrade changes how crisp your image looks. An audio upgrade changes whether people finish watching at all. Retention curves bend around intelligibility: when a viewer has to strain to follow a sentence, they leave.
The good news is that the parts of post-production sound that used to require a booth, a composer, and a rented studio are now largely software problems. Synthetic voice performance, generative music beds, and library-free sound effects can all be produced at a desk, iterated in minutes, and refined to a standard that holds up on phone speakers, laptop speakers, and a decent pair of headphones alike.
This guide is a practical workflow, not a tour of novelties. It covers what each audio layer does, how to generate and direct it, how to mix synthetic and recorded material together, and where the genuine limits still are. If you produce explainers, product videos, course content, documentary shorts, or social campaigns, the process below will replace a surprising amount of manual labour — and the last ten percent of polish is still, deliberately, your job.
Before generating anything, separate the soundtrack into layers. Mixing problems are almost always layering problems in disguise. If you cannot name each element's job, you cannot decide how loud it should be.
The spine of the video. Everything else exists to support it. Dialogue needs to sit forward, be free of room colour, and stay consistent in tone across cuts. Whether it was recorded in a booth, on location, or rendered from text, this layer sets the loudness reference for the whole mix.
Music carries emotion, pace, and section transitions. It also competes directly with speech. The most common beginner error is not "bad music" — it is music that is two or three decibels too loud, or music whose midrange sits exactly where consonants live. Generative music is extremely convenient here because you can regenerate a version that is less busy, or in a different key, in seconds.
This layer makes the picture believable. Footsteps, fabric movement, a keyboard click, the low hum of a room. Individually these sounds are barely audible; collectively their absence makes footage feel like it was shot in a vacuum. Ambience also masks edits and smooths the transitions between locations.
Not literally nothing, but deliberate breathing room. A half-second of true quiet before a key line is one of the cheapest and most effective tools in editing. Generative audio makes silence easy to preserve because you are not paying by the minute for a track that you will mostly mute.
Once these four layers are named, the workflow becomes orderly: generate or record dialogue, build a music bed that leaves space, place effects and ambience, then mix to a known loudness target.
Text-to-speech has crossed a threshold where the remaining variable is mostly the script and the direction, not the model. Here is a workflow that produces consistent results.
Read your script aloud before generating anything. Subordinate clauses that look elegant on a page collapse when spoken. Shorten sentences. Replace semicolons with full stops. If a sentence contains more than one idea, split it.
Synthetic voices handle punctuation and formatting poorly when they are inconsistent. Standardise units, expand ambiguous abbreviations, and write numbers the way you want them read if the voice mispronounces digits. Brand names with unusual pronunciation should be respelled phonetically for the generation pass, then corrected in the on-screen text so the viewer never sees the workaround.
Do not generate a thousand-word paragraph in one pass. Generate one to three sentences at a time. You get faster regeneration when a single phrase sounds wrong, better control of pacing, and cleaner edit points. Name your files by section so the assembly step stays sane.
Audition at least three voices reading the same two sentences. Listen for sibilance, how the voice handles commas, and whether the timbre matches the subject matter. A warm mid-range voice reads as trustworthy for a documentary; a brighter, faster read suits promotional content. Cast for the audience's expectation, not your personal favourite.
Lay takes on the timeline with 200–400 ms of silence between sentences. Longer gaps — around 600–800 ms — signal a new section. Consistent gaps are what make a generated read sound edited rather than produced.
Every voice model has a handful of words it mangles. Instead of regenerating the whole take, splice in a corrected word or short phrase from a second generation. Match levels at the join and crossfade by a few milliseconds.
The difference between a synthetic read that sounds acceptable and one that sounds professional usually comes down to direction. Treat the model like a performer who takes instructions literally.
Emotion: Most modern engines accept a style or emotion setting alongside the text. Use fewer, stronger choices. Switching emotion every sentence produces whiplash. Pick one emotional register per section — calm authority for the explanation, warmth for the conclusion — and hold it.
Emphasis: You steer emphasis with sentence structure, not with capital letters. Putting the key word last in a short sentence naturally lifts it. If a line must land hard, isolate it as its own take and adjust the surrounding silences rather than shouting at the model through punctuation.
Pacing: Aim for roughly 150–165 words per minute for explanatory content, faster for energetic promotional work. If the model offers a speed parameter, stay within about ten percent of natural speed; beyond that, artefacts appear in the consonants. For genuinely faster sections, regenerate at a slightly higher speed instead of time-stretching afterwards.
Intonation shape: Listen for the pitch contour at the end of sentences. A rising contour on a declarative sentence makes everything sound like a question. If you hear it, add a full stop, shorten the sentence, or try a different voice — some engines simply do this more than others.
Consistency: If your video is long, generate all voiceover in one session with one voice and one set of settings. Coming back three days later with a different style preset is how twenty-minute videos end up sounding like three different narrators.
A useful test: play the voiceover with your eyes closed. If you can follow the argument without pictures and without effort, the direction is working.
Generative music is most useful when you treat it as an adaptive scoring tool rather than a track download. Instead of hunting for one song that fits an entire video, you build short sections.
Decide your average shot length first. Fast-cutting sequences feel natural against 110–130 BPM; reflective sequences often sit better at 70–95 BPM. Generate a couple of options and cut picture to the music rather than forcing music onto a locked cut — it takes less time and looks more intentional.
Request an intro, a build, a main body, and an outro separately. That lets you drop the music out entirely under a technical explanation, then bring it back on a reveal. Full songs force you to fight their structure; sectional generation lets the structure follow your story.
Ask for arrangements that are sparse in the 1–4 kHz range, where speech intelligibility lives. Pads, soft piano, muted guitar, and light percussion work well. Dense synth leads and busy drum patterns will need heavy ducking, which makes the music sound like it is gasping.
If your tool can export stems — drums, bass, harmony, melody — take them. Muting the melody under dialogue and keeping the rhythm section gives you the energy of a full arrangement without the frequency collision.
For long videos, generate a two-to-four-minute bed and loop it with matched levels at the seam. Check the join with headphones; a small volume mismatch at a loop point is audible even when the musical content matches.
Effects work is where synthetic audio quietly outperforms stock libraries, because you can describe the exact object you need instead of browsing thousands of near-misses.
Generate short, dry sounds: a mug set down on a wooden table, a jacket zipper, a chair scuff, a page turn. Place them slightly before or after the visual action — a two-to-four frame lead is usually better than perfect sync, because it reads as intentional.
Every location has a room tone. Generate a low-level continuous bed for each distinct setting — office hum, street at night, forest edge — and run it under the whole scene at a very low level, roughly 20–30 dB below dialogue. This single step eliminates the "cut to silence" sensation that makes edited footage feel stitched.
Use these sparingly. A whoosh on every cut becomes noise. Reserve them for section changes, reveals, and scene transitions, and keep them short — under half a second for most formats.
Pan effects to match the picture. If the door closes on the right side of the frame, pan the sound right and add a touch more reverb to suggest the space behind it. Generative tools that accept reverb and distance parameters make this fast; if yours does not, add the reverb in the mix rather than regenerating.
Real recorded ambience plus a synthetic layer usually beats either alone. The recording provides authenticity; the synthetic layer fills gaps and covers edits.
Mixing synthetic material is not fundamentally different from mixing recorded material, but a few habits matter more.
Set your loudness target first. For most online video platforms, aim for roughly −14 LUFS integrated with a true peak ceiling of −1 dBTP. Podcast-style delivery often sits around −16 LUFS; broadcast specifications are stricter and typically around −23 LUFS. Pick the spec that matches your main distribution channel and mix to it once.
High-pass the voice. A gentle high-pass filter around 80–100 Hz removes rumble and plosive energy without touching presence. Synthetic voices sometimes carry more low-frequency energy than expected, so check before assuming it is clean.
Control sibilance. De-ess in the 5–8 kHz region if the letter S hisses. Use a narrow dynamic band rather than a broad static cut so you do not dull the whole voice.
Compress lightly. A ratio around 3:1 with a 10 ms attack and roughly 100 ms release keeps the read even without squashing it. Synthetic voiceover is often already quite consistent, so heavy compression usually just raises the noise floor.
Duck the music. Sidechain the music and ambience to the voice, targeting 12–18 dB of gain reduction under speech. Fast attack, slow release. The goal is for the music to feel continuous while never masking a syllable.
Check on small speakers. Phone speakers roll off below roughly 500 Hz. If your mix depends on bass for impact, it will vanish on the platform where most viewers actually watch. Balance so the voice and mid-range percussion carry the energy.
Limit, then listen again. A limiter at −1 dBTP catches stray peaks. After that, take a break, then listen to the whole piece once at normal volume without touching anything. Fatigue hides problems; fresh ears find them.
Music louder than the message. If you can hum the track but not repeat the last sentence, the balance is wrong. Pull music down 2 dB and re-evaluate.
One take for the whole script. Long single generations drift in tone and pacing. Generate in short sections and assemble.
Ignoring the gap between sentences. Uniform spacing is what separates "edited" from "generated." Vary gaps by section, not randomly.
Fighting the model on pronunciation. If a voice cannot say a term after three attempts with different spellings, record that one word yourself or choose a different voice. Time spent arguing with a synthesiser is time not spent editing.
No room tone. Cutting to absolute silence between scenes is the single clearest tell of an amateur mix.
Overusing effects. Every whoosh, riser, and impact you add competes with the previous one. If you can notice the sound design while watching, it is too loud.
Mixing on headphones only. Headphones reveal detail but hide balance problems. Always confirm on a laptop or phone speaker before exporting.
Forgetting the ending. Music that simply stops sounds like an error. Write or generate a short outro that resolves, even if it is only four bars.
Licensing is the least glamorous part of audio work and the one most likely to cause a problem after delivery. Establish a few habits.
First, know what your subscription or licence permits for the specific use: commercial advertising, client work, monetised publishing, and broadcast all have different terms. Second, keep a simple project record listing which voice, which music sections, and which effects were generated, along with the tool and the date. This is not bureaucracy — it is what lets you answer a client question six months later without guessing.
Third, be careful with voice cloning. A synthetic voice modelled on a real person requires that person's explicit, documented consent for the specific use. Do not clone a public figure, a former client's presenter, or a colleague for a pitch. Beyond the legal exposure, it destroys trust faster than any technical mistake.
Fourth, consider disclosure norms. Many platforms and audiences now expect synthetic narration to be indicated where it could be mistaken for a real person speaking. A short line in the description is usually enough, and it costs nothing.
Finally, deliver a clean package: the final mix at the agreed loudness, a version without music for clients who want to re-score later, and a dialogue-only stem. Clients who receive stems come back; clients who receive a single flattened file often do not.
Can synthetic voice carry an entire documentary? Yes, for narration-led formats. For interview-driven storytelling, recorded voices remain essential — the value is in the person, not the words.
How long does a ten-minute video take to sound-design? With a sectional workflow: roughly two to three hours for voice assembly and mixing, plus an hour for music and effects, assuming the script is final.
Should I generate music or use a library? Generate when you need exact tempo, length, or a sparse arrangement under dialogue. Use a library when you need a recognisable genre signature or a strong melodic hook.
What if the voice sounds robotic on long paragraphs? Shorten the paragraphs. Most "robotic" impressions come from unnatural sentence length and missing pauses, not from the model.
When should I record a real voice instead? Record live when the content depends on personality, when the speaker has authority the audience recognises, or when delivery must respond to something unscripted. Use synthesis for explanation, localisation, iteration, and volume.
Do I need a treated room for the recorded parts? Ideally yes, but a moving blanket behind the microphone and a quiet hour of the day solve most problems. Record 30 seconds of room tone and use it as your ambience bed.
How do I keep a series consistent? Save your voice, style, loudness target, and music-generation settings as a named preset, and document them in the project file. Consistency across episodes matters more than any single episode being perfect.
The broader point is that studio-quality sound is no longer a budget line. It is a workflow. Separate your layers, direct the performance, generate music in sections that respect the voice, place effects with restraint, mix to a known target, and check the result on the worst speaker your audience actually uses. Do that consistently and your videos will feel finished in a way that a better camera never achieves.

Alexander
Learn how to turn a written idea into a polished short film with AI video tools: scripting, model choice, shot consistency, sound design, and editing.
Learn how AI transcription enables text-based editing, faster reviews, accessibility, and reliable post-production workflows from rough cut to final delivery.
Learn a repeatable workflow for photorealistic AI video: model selection, shot planning, prompt design, continuity, audio, and quality control.