The Three Things That Change After You Hit Upload
- ● Your master file is never touched
- ● Every upload gets re-encoded
- ● Playback loudness is normalised
The short answer is that nothing happened to your file. Three things happen to the copy the platform serves, and the fix for each is different, so choose the one that matches what you are hearing.
The platform re-encodes your audio, it resamples if your rate does not match its target of 48 kHz, and it normalises playback loudness. If your voiceover sounds thinner rather than muffled, the loudness stage is the usual answer.
The rest of this guide separates the three, because the common advice to just export at a higher bitrate only addresses one of them.
The Sample Rate Mismatch Nobody Set Deliberately
Most AI voice tools deliver 44.1 kHz audio. ElevenLabs, for example, exports at 44.1 kHz across its plans, with the bitrate rather than the sample rate changing between tiers.
Video platforms want 48 kHz. YouTube publishes 48 kHz as its recommended upload sample rate, which means a 44.1 kHz voice track gets converted somewhere along the way.
That conversion is not automatically bad. It becomes audible when it happens late, in a hurried encoder pass, on top of a file that has already been compressed once.
The cheap fix is to convert once, deliberately, in your editor. Set the project to 48 kHz before you drop in the narration, and the resample happens in a controlled place instead of at the end of a chain.
Two Rounds of Compression Do the Damage
Your voice file is usually lossy before you ever open the editor. A 128 kbps export is a compressed representation, and the encoder threw away detail it decided you would not miss.
Then the platform compresses again from your video upload. The second encoder is working on the artefacts of the first, and this is where consonants get smeared and room tone starts to swirl.
This is why the highest-quality export you can get matters more than the bitrate you upload at. On ElevenLabs, 128 kbps is the ceiling on Free and Starter, while Creator and above list both 128 and 192 kbps, as of September 2026.
If your plan offers an uncompressed or PCM output through the API, that is the one to use for narration you intend to edit. ElevenLabs lists 44.1 kHz PCM output via the API on its Pro plan, priced at $99 per month as of September 2026. Confirm current plan limits on the official pricing page, since these tiers move.
Loudness Normalisation Undoes Your Mastering
The most common complaint is that the voice sounds quieter and flatter online than in the editing timeline. That is normalisation, not degradation.
Streaming platforms turn loud uploads down to a consistent playback level so viewers are not adjusting volume between videos. Independent measurements put the typical target near minus fourteen LUFS, though YouTube does not publish the figure.
The practical consequence is counterintuitive. Pushing a track hard with a limiter does not make it louder for the viewer, it just means the platform reduces it more, and you keep the squashed dynamics without the volume.
So master for tone rather than for level. A voiceover with intact dynamics that sits below the normalisation target sounds better after the platform is done with it than one crushed to be as loud as possible.
What Your Editor Does In Between
Video editors are a quiet source of quality loss because their defaults were chosen for convenience. A project set to 44.1 kHz with a 48 kHz timeline will resample every clip on export.
Track effects add another layer. A normalisation plugin left on a bus, or auto-ducking against background music, changes the narration in ways that only show up when you listen to the finished render rather than the timeline.
Background music is where most of the damage is done. Voice sits in the same frequency band as most music beds, and a bed mixed only a few decibels down eats the clarity you paid a voice tool for.
The audible test is simple. Render thirty seconds with music and thirty seconds without, then compare the narration in both, a habit that also helps when tuning delivery as covered in AI voice pacing and pauses.
The Numbers Worth Matching Before Export
- ● 48 kHz matches the platform target
- ● 384 kbps stereo is the published guidance
- ● Voice tools often cap at 128 kbps
The table below pairs the published upload guidance against what voice tools typically hand you, so you can see where the gaps are in your own chain.
| Setting | Platform guidance | What voice tools give you | What to do about it |
|---|---|---|---|
| Container | MP4 | Not applicable, audio only | Set the editor export, not the voice tool |
| Audio codec | AAC-LC or Opus | MP3 by default | Keep MP3 out of the final render if you can |
| Sample rate | 48 kHz | 44.1 kHz on most plans | Convert once in the editor project settings |
| Stereo bitrate | 384 kbps | 128 kbps, or 192 kbps on higher tiers | Export the video track well above the voice source |
| Mono bitrate | 128 kbps | Same range | Narration in mono is fine and saves bandwidth |
| 5.1 bitrate | 512 kbps | Not offered | Irrelevant for talking-head or narration work |
Read the sample rate row first, since it is the one people never look at. Everything else in the table is a headroom question, and headroom is cheap.
The bitrate rows also explain a frequent misunderstanding. Uploading at 384 kbps does not repair a 128 kbps source, it only avoids adding a third round of loss on top of it.
How to Tell Which of the Three You Are Hearing
The three causes leave different fingerprints, and five minutes of listening beats guessing at settings.
Download your own published video and play it against your master file. If the published version is simply quieter and less lively, that is normalisation, and no export change will alter it.
If consonants sound smeared and there is a faint swirl behind the voice, you are hearing repeated compression. That points at the source export quality and at how many apps the file passed through.
A brittle, slightly glassy tone on sibilants usually means a late resample. Check whether the voice tool, the editor project and the export all agree on a sample rate, because one disagreement is enough.
Which Fix Fits Your Setup
- ● Podcasters should fix loudness first
- ● Course creators should fix sample rate
- ● Short-form editors should fix the codec chain
Podcaster publishing audio only: Loudness is your problem, not bitrate. Stop limiting the narration hard, leave the dynamics intact, and let each platform normalise from a clean master.
Course creator with long narrated modules: Fix the sample rate first, because a resample repeated across dozens of lessons is the most consistent quality loss in that workflow. Set the project to 48 kHz once and every module inherits it.
Short-form editor cutting fast talking videos: Watch the codec chain. Short clips often pass through three apps before upload, and each hop re-encodes, so export uncompressed between tools even though the files are larger.
Anyone using the cheapest voice plan: Your 128 kbps ceiling is the binding constraint, and no export setting fixes it. Judge whether the upgrade is worth it by comparing the same script at both bitrates before paying, using the comparison in best AI voice generators to see which tools expose higher-quality output.
Audiobook or long narration work: Uncompressed source is not optional, since distributors have their own encoding requirements and will reject or re-encode weak files. The trade-offs are set out in AI narration vs a voice actor for audiobooks.
Anyone mixing narration with music beds: Solve the mix before touching export settings. No codec choice rescues a voice buried under a bed, and the fix costs nothing but attention.
The Export Recipe to Save Once and Reuse
Set your editing project to 48 kHz before importing anything. This single step removes the most common avoidable resample from the whole chain.
Pull the highest-quality output your voice plan allows, and keep that file as the master rather than the version you dropped on a timeline. Re-rendering from a compressed intermediate is what turns one round of loss into three.
Keep one reference clip from a video you were happy with. Comparing a new render against a known good one catches a drifting setting faster than any checklist, and it takes about a minute.
Then leave the loudness alone. Mix so the narration is clearly above the music, keep some dynamic range, and let the platform do the levelling it is going to do regardless.
Save those settings as an export preset and stop thinking about it. If you are still choosing a tool for regular video work, our guide to AI voice generators for YouTube videos covers which output options matter for this workflow.
FAQ
Did the platform ruin my audio or was it bad before upload?
Almost always no. The file on your drive is untouched, and what changed is the platform transcode plus loudness normalisation applied at playback. Download your own upload and compare it with the master to hear the difference for yourself.
Does exporting at 48 kHz really make an audible difference?
Not by itself, but it removes one avoidable resample. YouTube asks for 48 kHz, while most voice tools export at 44.1 kHz, so converting once during the edit is cleaner than letting the encoder do it late.
Is a higher bitrate always better for the upload?
Only up to a point. Uploading a heavily compressed 128 kbps voice file guarantees a second lossy pass, but going far beyond the recommended 384 kbps stereo mainly costs you upload time.
Should I still normalise my own audio if the platform does it anyway?
Yes, and that is the usual cause of a voiceover sounding flat after publishing. Platforms normalise loudness at playback, so a track pushed hard in mastering gets turned down and loses the punch it had locally.
Why does my voice tool only offer one export quality?
Mostly a limitation of the plan. The ElevenLabs Free and Starter tiers export at 128 kbps and 44.1 kHz, while Creator and above add 192 kbps, so the ceiling is commercial rather than technical.
Sources
- YouTube Help: Recommended upload encoding settings — checked 2026-10-02
- WHO: Making Listening Safe — checked 2026-10-02
- ElevenLabs: Pricing — checked 2026-10-07
Some links may be affiliate links. We may earn a commission at no extra cost to you.
How this article was made: it was drafted with AI assistance from the official sources listed above, and its figures were checked against those pages on the dates shown. It is not based on personal hands-on testing.
Comments
Post a Comment