How to Make a YouTube Voiceover with ElevenLabs
A practical 2026 path: write a 2–5 minute script, pick a Voice Library voice (or reuse a clone if the channel is already your voice), generate the read in Text to Speech — in chunks if the draft is long — download the file, and drop it on the video timeline. Built from ElevenLabs' live Text to Speech help and pricing (August 2026).
Most “AI YouTube voiceover” posts skip the two facts that actually decide the file. First: a two-to-five minute narration is not a fifteen-second sting, and it is not a clone training set. Second: a free-tier preview is not a track you put under AdSense. This guide is the working path for the spoken episode — script, stock voice, generate, download, drop it on the video — using the one voice desk we send for this brief: ElevenLabs.
When a stock ElevenLabs voice is enough — and when you clone
Use a Voice Library or Default Voice when the channel has no existing vocal identity. A retention-graph explainer. A product teardown on B-roll. A faceless weekly desk that is not a named host. The economic case is the recut: edit the sentence that went stale, generate again, replace the clip. You do not re-book a booth for four minutes.
Clone only if viewers would notice a stranger. Commentary, a host who already talks on camera and wants the same timbre on B-roll weeks, a personal-brand channel that is one voice — that is a training-set and rights job, and it lives at How to clone your voice for YouTube. Skip both if you already record the episode yourself and the words will not change. That file is a recording, then a cut. Skip the stock library if the picture already exists on slides and the missing piece is a directed studio narrator. That is How to make training voiceovers with Murf. Murf is the foil, not a Try button.
The seven-step ElevenLabs YouTube voiceover
- 01
Write a 2–5 minute YouTube script — hook, promise, two or three beats, close
A YouTube voiceover is a spoken episode, not a sting and not a first-second Reel hook. Time it out loud at a normal pace: two minutes is a tight explainer; five is a full beat-by-beat walkthrough. Open with the claim that earns the next sentence, name what the viewer can do after the watch, then walk two or three proofs. Close with one recap and one next click on this channel. Spell numbers, product names, and abbreviations the way they should be heard. If you wanted a 10–20s show open, that is the podcast-intro how-to. If you wanted a 15–60s public hook, that is the TikTok / Reels voiceover how-to.
- 02
Confirm a generated VO is the right format
Use ElevenLabs when the missing piece is a realistic narration you can regenerate when a fact or a title changes. Skip it if you already record the episode in your own voice and the words will not be rewritten — film or record that, then cut it. Skip it if the job is a longer directed narrator synced to slides. That is Murf Studio, and that how-to is an ordinary path on this site, not a hop here. Skip a library voice if the channel is your voice and viewers would notice a stranger. That is a clone, and the full training-set path lives on the YouTube clone how-to.
- 03
Open Text to Speech — the speech playground, not Agents
ElevenLabs’ published product for this brief is Text to Speech: paste text, pick a voice, generate speech, download a file. That is the desk this page is written for. ElevenAgents is a different product (conversational agents). Studio is a longer project surface for chaptered work. We are not inventing a third workspace. This page sends you to ElevenLabs only.
- 04
Sit on a plan that grants commercial rights before you treat the file as the upload
ElevenLabs’ own docs are explicit: you keep ownership of generated audio, but commercial usage rights come with paid plans. Free is an audition — published pricing lists 10,000 credits a month and no commercial license. Starter is the first paid tier that lists a commercial license and Instant Voice Cloning. Professional Voice Cloning is listed on Creator and above. Confirm the live grid on elevenlabs.io/pricing. We do not invent a checkout total or a commission rate. Do not publish a free-tier narration on a monetized video.
- 05
Pick a Voice Library voice — clone only if the channel already is your voice
Open Voices and browse Default Voices or the Voice Library. Preview before you apply. Cast one narrator and keep it for the episode and for later recuts. Filter toward narration-style reads, then listen on headphones. A stock voice is enough when the channel has no existing vocal identity — listicles, explainers, B-roll essays, a faceless desk that is not a named host. Clone only if viewers would notice a stranger. Instant Voice Cloning is the published self-serve path from short samples (ElevenLabs’ cloning help: about 1–2 minutes of clean audio) on paid plans that list it. Professional Voice Cloning trains a dedicated model on a longer set (published as 30–180 minutes) and requires Creator or above. Do not invent a clone wizard here — if you need the training-set workflow, use the clone how-to. Do not clone a guest, a commenter, or a voice you found on someone else’s channel.
- 06
Paste in sections, pick a model, generate, then re-roll the weak beat
A typical 2–5 minute English script is a few thousand characters. Multilingual v2 is the published “most stable on long-form” model and lists a 10,000-character cap — most working YouTube VOs fit in one paste. Eleven v3 is the expressive model (5,000-character cap, audio tags such as [sighs] or [clears throat]); a dense five-minute draft can hit that wall. ElevenLabs’ own Text to Speech help for large conversions: split long text into segments. Generate the hook as its own take, then each beat, then the close — even when the whole script would fit — so a flubbed open does not burn a clean body. Voice first, then model, then settings — that is the order ElevenLabs’ Text to Speech guide ranks. Official starting settings for the sliders that exist: Stability around 50, Similarity around 75, Style exaggeration at 0. Speed defaults to 1.0 (range 0.7–1.2) where the model offers it. Spell out numbers. Then press Generate Speech. The model is nondeterministic — same text can yield a different take. Re-roll the weak section; leave a clean section alone.
- 07
Download the file, then drop it on the video timeline
After a generation, ElevenLabs’ help says you can download immediately from the control on the bottom right. Older takes live in History on the Text to Speech page — History lists MP3 (128 kbps) or WAV, with Advanced formats of MP3 192 / 256 kbps, M4A, and FLAC. Higher-quality options are listed on paid tiers; confirm the live download list. Name each chunk so the assembly is obvious:
retention-graph_hook_en_v1,retention-graph_beat2_en_v1. Then put the files on the video: import the WAV or high-bitrate MP3 onto a dedicated VO track, line the hook to picture, leave a breath between sections, and duck any music bed under the narration. When a fact changes, edit that section, generate, bump to v2, replace the clip. That recut is why you did not book a booth for a four-minute explainer.
Stock VO vs clone vs Murf narrator
“AI YouTube voiceover” is a search, not a product. The decision is the format. The ElevenLabs column matches Text to Speech help and the live pricing page as of August 2026. Murf and the clone path stay qualitative here — we already wrote those desks, and this page does not hop there.
| Criterion | ElevenLabs stock VO | Clone the host | Murf studio narrator |
|---|---|---|---|
| What the viewer hears | A realistic 2–5 min narration you can regenerate when a fact or a title changes | The named host, from a training set — Instant or Professional, not a library voice | A directed studio narrator timed on a timeline — built for lessons, usable as a read |
| When it is the right buy | The episode needs a human-sounding VO and the channel is not already a named voice | Viewers would notice a stranger — commentary, personal-brand explainers | The picture already exists — slides or a screen recording — and you need a timed studio read |
| What you re-do when the copy changes | Edit the section, Generate Speech, replace the clip on the timeline | Same generate-and-replace once the clone exists — the cost was the training set | Edit the block, regenerate, replace the audio |
| Tool this desk writes for | ElevenLabs Text to Speech — script, Voice Library, Generate Speech, download | Ordinary path: the YouTube clone how-to. Not a hop on this page | Murf Studio — ordinary path: the training-VO how-to. Not a hop on this page |
| Best 2026 fit | Faceless or explainer YouTube that should sound like a person, not a stock LMS read | A channel that already is the host’s voice — see the clone how-to | A module narrator you time to slides — see the Murf training page |
ElevenLabs
A Text to Speech narration for a video that already has picture. Free plan to audition a script; a paid plan is the commercial-rights download you can put on YouTube.
The script is a 2–5 minute episode, spoken
Do not write a cold open for a documentary. Do not write a first-second Reel hook and then pad it. Time the copy out loud. Two minutes is enough for one claim and two proofs. Five minutes is a walkthrough. Longer than that and you are writing a different video.
A working shape, spoken at a normal pace — roughly two and a half minutes:
Your retention graph is not a thumbnail problem. The first cliff at eight seconds is the sentence you wrote after the title, not the picture you chose for the still.
I’m going to walk the three drops most explainer channels actually have: the open, the first proof, and the mid-roll stall. You will leave with one rewrite for each.
Drop one: you restated the title. The viewer already clicked. Say the consequence instead. “If this line is wrong, your next four minutes are a comment, not a watch.”
Drop two: you promised three tips and started with a definition. Put the first proof in the first forty seconds. A number they can check in YouTube Studio beats a glossary.
Drop three: you changed topics without a breath. End a beat on the action — “open Advanced mode, not Overview” — then start the next beat. Do not announce the section.
Rewrite those three lines. Re-export the VO. Watch the graph on the next upload before you buy another thumbnail pack.
Spell the words the model should say. “Eight seconds,” not “8s.” “YouTube Studio,” not “YT.” ElevenLabs’ Text to Speech help is blunt about numbers and symbols: write them out, especially on multilingual models, because the same digit is pronounced differently across languages. A video about “Q3 ARR” needs those letters in the script the way you want them heard.
Keep one voice for the episode and for the recut. A library narrator that survives your product name is worth more than a cinematic whisper that flubs the title. Emotional range is not why you are here. If realism-versus-timeline is the actual question, the comparison is ElevenLabs vs Murf vs Synthesys — ordinary path, no hop. Full product notes live in our ElevenLabs review.
Pick a voice. Generate in chunks. Direct the take.
Open Text to Speech. Select a voice from the control ElevenLabs documents at the bottom left — Default Voices or the Voice Library. Preview. Apply one voice and keep it. A viewer who hears January’s explainer and October’s recut should still recognize the person who said the title.
Clone only if that person has to be you. Instant Voice Cloning is the published fast path from short samples (about 1–2 minutes of clean audio in the cloning help). Professional Voice Cloning is the dedicated model: Creator plan or above, a longer training set (published as 30–180 minutes), and a wait while it fine-tunes. We are not going to invent a clone wizard on this page. If you do not already have the voice in My Voices, use How to clone your voice for YouTube and come back.
Paste in sections, not one unbroken essay. Voice, then model, then settings — ElevenLabs ranks those in that order. Multilingual v2 is the published stable default and the one we would start on for a 2–5 minute English VO. It lists a 10,000-character cap; most working scripts fit in one generation. Still split the hook from the body. A clean two-minute middle is not worth burning to fix the first twelve words. ElevenLabs’ own help for large conversions is the same instruction: split long text into segments.
Eleven v3 is the expressive model: audio tags such as
[sighs] or [clears throat], a
5,000-character cap, and fewer of the older sliders (Speed,
Similarity, and Speaker Boost are documented as unavailable
on v3). A dense five-minute draft can hit that wall — another
reason to generate beat by beat. Flash models are the
low-latency family; a YouTube preroll does not need 75ms.
Where the sliders exist, the official starting point is Stability around 50, Similarity around 75, Style exaggeration at 0. Speed defaults to 1.0; the documented range is 0.7 to 1.2. Lower Stability for a livelier take, then generate more than once — the model is nondeterministic. Higher Stability for a straighter explainer. Do not invent a slider we cannot see on the public page. If a control is missing on your model, change the sentence and re-roll.
Pauses: punctuation first. A dash or em-dash is the documented beat; ellipsis adds hesitation, which a YouTube open usually does not want. On Multilingual v2, Flash v2, and Flash v2.5, ElevenLabs also documents an SSML break tag for a timed pause of up to three seconds. Confirm the live syntax in their Text to Speech help rather than pasting markup we cannot see unchanged. On Eleven v3, use audio tags and punctuation — that model’s help says it does not support SSML break tags.
Press Generate Speech. Listen on headphones. Re-roll the hook until the first sentence lands. Leave a clean beat alone. Two free regenerations of the exact same text and settings are the published allowance; any edit to the copy or the sliders is a new generation.
Download, then put the file on the video you already have
The download is the gate. ElevenLabs’ Text to Speech help: after you generate, use the download control on the bottom right. Earlier takes sit in History on the same page — sidebar Text to Speech, then the history panel (or the history icon above Generate Speech on a narrow screen). History lists MP3 at 128 kbps or WAV; Advanced adds MP3 192 / 256 kbps, M4A, and FLAC. Higher-quality options are listed on paid tiers. For a YouTube handoff, WAV or a high-bitrate MP3 is the usual file. Confirm the live list on the plan you pay for.
Then hand the file to the video you already have. Import each
chunk onto a dedicated VO track. Line the hook to the first
picture change. If you generated in sections, leave a breath
between clips so the join does not click. If you use a music
bed, duck it under the narration so the first sentence is the
loudest thing in the open. Name the export so the next person
can find it: retention-graph_hook_en_v1. When a
number changes, open that section, generate, bump to v2,
replace the clip. Leave the beats that are still true. That
is the reason you did not film the host for the VO.
Viewers still deserve a plain-language note when the narration is synthetic. Say it in the description. Use YouTube Studio’s current checkbox for realistic synthetic or altered content on upload. That is separate from the vendor license. You need both.
Commercial rights, the upload, and the next cut
Read the live page. ElevenLabs meters in character credits across the suite. Free is 10,000 credits a month and no commercial license. Starter is the first paid tier that lists a commercial license and Instant Voice Cloning — $6/month on the public grid we verified in August 2026 (elevenlabs.io/pricing). Creator is the tier that lists Professional Voice Cloning ($22/month on that same grid; annual billing is cheaper). Confirm checkout. We do not print a commission rate.
A weekly 2–5 minute VO is the plan question, not a rounding error. Sit on a paid plan before the file is the one under ads. If you later want the host’s likeness on B-roll episodes, that is the clone how-to, not a reason to start this brief in Creator. If the remaining job is a slide narrator, the Murf training page is that brief. If the remaining job is a fifteen-second open, the podcast-intro how-to is that brief. The voice-tool comparison lives on ElevenLabs vs Murf vs Synthesys. This page does not hop there. The only Try button here is ElevenLabs.