Why audio decides whether your video gets watched
A viewer will forgive a slightly soft shot. They rarely forgive muddy narration or a music bed that fights the voice. Speech carries the information payload of a tutorial, a product demo, or an explainer, and music sets the emotional expectation for everything on screen. When either layer is wrong, the viewer spends attention decoding instead of absorbing, and the exit happens earlier.
Retention curves tend to follow the same shape across niches: a drop in the first fifteen seconds, then a slow decay. Poor audio accelerates both. A hum in the background, a level jump between two takes, a music bed that ducks clumsily under every sentence — these are small problems individually and a reason to click away collectively.
The traditional fix was a treated room, a decent microphone, and an afternoon of retakes. That still produces the best results when you have the time. The modern fix is a hybrid pipeline: write for the ear, record or generate the narration, generate or licence the music, then mix both to consistent loudness targets. AI changes the cost and the iteration speed of that pipeline, not the fundamentals. You still need a clean script, a deliberate voice choice, and a mix that puts the message first.
Think of your audio as a three-part stack: narration, music, and the mix that binds them. This guide walks through each part in the order you will actually work — script, voice, music, mix, delivery — and flags where automation helps and where it quietly makes things worse.
Narration and music are two different jobs
Treating narration and music as one "sound" problem is the most common source of rework. They have different jobs, different failure modes, and different quality bars.
Narration: clarity first, personality second
Narration exists to deliver information with as little friction as possible. The quality bar is intelligibility: every word recognisable on a phone speaker at half volume. Personality is a modifier, not the foundation. A warm, characterful read that mumbles is worse than a neutral read that is clear.
Three variables control intelligibility in practice: pace, articulation, and dynamic consistency. Pace sits around 140–165 words per minute for most explainer content; faster works for energetic social edits, slower for technical instruction. Articulation is about consonant endings — the "t" in "content" and the "s" in "risks" — which is exactly where cheap synthesis fails. Dynamic consistency means the voice does not leap in volume between sentences, which happens when you stitch together takes recorded in different rooms or generate lines in separate sessions with different settings.
Music: continuity, energy, and space
Music does three jobs in a video: it smooths hard cuts, it sets and holds an energy level, and it tells the viewer when a section has ended. It should almost never be the thing the viewer notices. If someone remembers your background track more than your point, the balance is wrong.
A practical way to think about it is "emotional bandwidth". A dense, melodic track occupies the same attention space as a voice, so it competes. A sparse, textural track — pads, light percussion, sustained low notes — leaves room for speech while still doing continuity work. That is why so much documentary and corporate music sounds minimal: it is designed to be ignored, on purpose.
Where the two layers collide
Narration and music fight in one specific region, roughly 200 Hz to 4 kHz, which is where speech intelligibility lives. Music that is busy in that band will bury a voice no matter how much you lower the overall level. The fix is not always "turn the music down"; often it is "choose music that is not doing much in the midrange while the voice is talking".
A voiceover workflow that scales
Step 1 — Write for the ear
Read every line aloud before you record or generate it. Sentences that look fine on a page often collapse when spoken: nested clauses, stacked numbers, and parenthetical asides. Aim for one idea per sentence, twelve to eighteen words, active voice. Write numbers the way you want them read ("twelve", not "12") if your tool handles pronunciation literally. Spell out acronyms on first use, and prepare a pronunciation guide for brand names, product names, and place names.
Step 2 — Choose a voice and commit
Pick a voice profile based on three criteria: register, pace, and accent fit. Register means pitch and timbre — lower registers read as authoritative, mid registers as friendly, higher registers as energetic. Pace is how the model or presenter naturally flows; match it to your content type. Accent fit is about your audience, not your preference. A neutral accent travels further across markets; a regionally specific accent builds trust in that region.
Commit to one voice across a series. Consistency compounds: viewers recognise the voice before they read the title. If you need a second voice for dialogue, differentiate it clearly in register rather than slightly in tone.
Step 3 — Control pacing, pauses, and pronunciation
Pauses are punctuation you can hear. Add a short pause (250–400 ms) between sentences, a longer one (600–900 ms) after a section change or before a key number, and a breath where a human would take one. Most text-to-speech interfaces accept punctuation, break markers, or SSML-style tags; even simple comma and ellipsis tricks can shape a read meaningfully.
Stress is the other lever. If a sentence carries a contrast — "this is not the cheap option, it is the fast one" — mark the emphasised words. Slight pauses before and after an emphasised word do more for comprehension than raising volume.
Step 4 — Review by ear, not by waveform
Stop listening to individual lines and listen to the whole track at 1.5x speed. Speed listening exposes awkward transitions, repeated sentence rhythms, and pitch drift that normal-speed listening hides. Then listen once on a phone speaker in a noisy room. That is the environment most of your audience is in.
Generating background music that follows the edit
Match tempo to cutting rhythm
If your edit cuts roughly every two seconds, a track at 120 BPM gives you a beat every half second — four beats per cut. That feels steady. A track at 90 BPM gives you a beat every 0.67 seconds, which pushes against the cut rhythm and can feel restless. The practical rule: decide your average cut interval, then choose a tempo that divides evenly into it, or a tempo exactly half or double that.
Build in segments
Long videos rarely work with one continuous bed. Split the audio into segments that match your structure: a short intro sting, a body bed at lower intensity, a bridge or transition motif, and an outro that resolves. Generating each segment separately gives you control over energy and makes transitions deliberate rather than accidental.
For a ten-minute explainer, a workable structure is: a four-second intro sting, three minutes of body bed, a two-second transition, another four minutes of body bed at a slightly different texture, then an outro that resolves the harmony. Save each segment as a separate file so you can re-time them without regenerating everything.
Leave room for the voice
Generate or select music that is sparse in the 300 Hz–3 kHz range, or apply a gentle EQ dip there. If your generator supports stems, export with drums and bass separated so you can duck only the melodic elements under speech and keep percussion steady. A steady rhythmic pulse under a talking head keeps energy up without competing for intelligibility.
Mixing and mastering for how platforms actually play audio
Set a loudness target and stick to it
Streaming platforms normalise loudness, typically somewhere near -14 LUFS integrated for music-forward services, with true peak ceilings close to -1 dBTP. If you deliver much louder, the platform turns you down and your dynamics get squashed relative to everything else. If you deliver much quieter, you sound weak next to competitors.
A practical setup: target -14 to -16 LUFS integrated for narration-led video, with narration peaking around -6 dBFS and music sitting 15–20 dB below the voice when speech is present.
Duck without pumping
Sidechain compression lowers music automatically when the voice is active. Done badly it creates an audible "pumping" effect where the music breathes in and out. Three fixes: slow the attack to 20–40 ms so the duck is not instant, set the release to 300–600 ms so it recovers smoothly, and reduce the duck amount to 4–8 dB instead of 15 dB. If you can hear the ducking working, it is too aggressive.
Carve frequencies instead of only lowering volume
A high-pass filter at 80–100 Hz on the voice removes rumble without thinning it. A narrow dip of 2–4 dB in the music around 1–3 kHz creates space for consonants. This combination lets you keep music at a satisfying level while the voice stays clear.
Test the phone speaker
Mix on monitors, then check on a phone, then on cheap earbuds. Phone speakers roll off below roughly 500 Hz, so if your mix depends on low-end warmth for clarity, it will sound thin and muddy. If the narration is intelligible on a phone speaker, it will be intelligible anywhere.
Choosing tools: decision criteria that matter
Text-to-speech engines
Evaluate on four axes: naturalness, control, language coverage, and export flexibility. Naturalness you can test in ten minutes with a paragraph containing numbers, acronyms, and a question. Control is about how finely you can adjust pace, emphasis, and pauses without re-recording — this matters more than raw voice quality for long-form work. Language coverage matters if you localise; check that the same voice exists across the languages you need, because switching voices between language versions breaks series identity. Export flexibility means WAV output without added processing, so you can mix properly.
Music generators
Evaluate on structure control, stem export, and licence clarity. Structure control means specifying duration, intensity curve, and whether the track should build or stay flat. Stem export lets you duck selectively. Licence clarity is non-negotiable: you need to know whether you can monetise, whether attribution is required, and whether the track can be used in client work.
Voice cloning, consent, and disclosure
Cloning your own voice is straightforward and useful for pickups and corrections. Cloning someone else's requires documented permission, ideally with a defined scope: which projects, which duration, which markets. Many platforms require disclosure when synthetic voice is used in certain categories, and audience trust erodes fast when synthetic narration is passed off as a real interview or a real testimonial. Keep a simple record: whose voice, what permission, what the synthetic audio was used for.
A worked example: a six-minute explainer
Take a product explainer with five sections. Here is the sequence that produces a publishable file in about ninety minutes.
- Script pass (20 min). Read aloud, cut nested clauses, mark pauses and emphasised words, write out numbers and acronyms.
- Voice generation (10 min). One voice profile, consistent settings, generate section by section for easier retakes. Export WAV.
- Music generation (15 min). One intro sting, one body bed, one transition, one outro. Export stems where possible.
- Assembly (15 min). Lay narration on track one. Place music segments aligned to section boundaries, not to the start of the timeline.
- Mix (20 min). High-pass the voice, dip the music in the midrange, apply sidechain ducking at 5 dB, set integrated loudness around -15 LUFS.
- Check (10 min). Phone speaker pass, 1.5x speed pass, waveform glance for clipping.
The most common failure in this workflow is generating all the narration first and discovering during assembly that section lengths have changed because the music demands a different pacing. Build the picture edit first, lock timing, then generate audio to the locked edit.
Common mistakes and how to fix them
Identical sentence rhythm. If every sentence has the same length and cadence, the narration becomes hypnotic in the wrong way. Vary sentence length deliberately: a long explanatory sentence followed by a short one.
Music with vocals under narration. Sung vocals compete directly with speech. Use instrumental beds under dialogue unless you are deliberately creating a montage moment with no narration.
Full-volume music in the intro and outro. These are the moments where the music is uncovered and can be loudest. In the body, with speech present, pull it down.
One long music file for a long video. Repetition becomes noticeable after about ninety seconds. Either generate two or three variations or use a track with clear structural sections.
No headroom. If you normalise everything to 0 dBFS, you leave no space for the final limiter and risk distortion after platform encoding. Leave 1 dBTP of headroom.
Ignoring breath and mouth noise. A generated voice with no breath sounds uncanny over long durations. A breath inserted every four to six sentences restores naturalness.
Skipping the phone test. Everything can sound fine on good headphones and fail in the real world.
Rights, disclosure, and a fast quality checklist
Rights questions usually reduce to three: can I monetise, must I attribute, and can I use this in client work. Read the licence for the specific asset, not the marketing page. Keep a folder per project with the source, the licence type, and the date.
Disclosure is a trust issue more than a legal one. If the voice is synthetic, a small note in the description is enough for most audiences and protects you if a viewer later notices.
Checklist before publishing:
- Narration intelligible on a phone speaker
- Loudness around -14 to -16 LUFS integrated, peaks below -1 dBTP
- Music ducked 15–20 dB under speech, not audibly pumping
- Music segments aligned to section boundaries
- Pronunciation verified for all names, numbers, and acronyms
- Licences stored per project
- Disclosure added where required
FAQ
How long should an intro music sting be? Two to four seconds. Long enough to register, short enough that it does not delay the content.
Can I mix generated and recorded narration? Yes, but match the room and the processing. Recorded voice usually needs a light high-pass and gentle compression to sit with generated lines.
What if my voice generator mispronounces a word? Respell it phonetically in the script, or use a pronunciation dictionary if the tool supports one. Never fix it by cutting and pasting from another take with a different setting.
Does a faster read hurt retention? Only if intelligibility drops. Slightly faster than comfortable usually improves completion rates for social edits; technical content should stay slower.
Should I use the same music across a series? A recognisable intro sting builds identity. The body bed should vary to avoid fatigue.
How much does the mix matter compared to voice quality? A mediocre voice with a good mix outperforms a great voice with a bad mix, because intelligibility is a mix property as much as a recording property.
