Eleven v4 for Tutorial Narration: What Changed (2026)
Vorec Team · 2026-10-19 · About 9 min read
If you have ever generated a voiceover for a software tutorial, you know where it usually goes wrong. The words are right, but every sentence lands with the same weight. A warning ("don't click Delete yet") sounds exactly like a menu label. Your product's name comes out mispronounced in step three. And when you regenerate one line, the new take doesn't quite match the voice around it.
On 28 September 2026 ElevenLabs released Eleven v4, a text-to-speech model, and a low-latency variant called Eleven v4 Turbo. Several of the changes ElevenLabs describes address exactly those problems: directing how a line is delivered, more reliable custom pronunciation, steadier voices across regenerated lines, and voices that keep their identity in other languages.
This post sets out what ElevenLabs announced and then what it means for people who narrate screen recordings, product walkthroughs and training videos. Performance figures are ElevenLabs' own claims, and we label them that way. We have not tested the models.
Checked on 19 October 2026. Source: the ElevenLabs blog post Introducing Eleven v4, our most emotive model, published 28 September 2026. On the day we checked, the page showed a "last updated" date of 19 October 2026; its text matched the copy we had saved earlier. Model behaviour and availability can change, so confirm details in ElevenLabs' own documentation before you plan around them.
What ElevenLabs released
| Model | What ElevenLabs says it is for | Where it is available (per ElevenLabs) |
|---|---|---|
| Eleven v4 | Its most emotive text-to-speech model so far, built on a new architecture | ElevenAgents, ElevenCreative and ElevenAPI |
| Eleven v4 Turbo | The same technology tuned for low-latency uses such as voice agents | ElevenAgents, ElevenCreative and ElevenAPI |
Other points from the announcement:
- Languages. ElevenLabs says both models support more than 90 languages.
- Voice cloning. ElevenLabs says Instant Voice Clones can now capture a voice well from about 10 seconds of audio, and that Eleven v4 adds support for Professional Voice Clones.
- Free account access. The post invites readers to sign up for a free account to try either model. It doesn't list per-plan limits for v4. Our ElevenLabs pricing breakdown covers the plans themselves.
The benchmark claims, with their conditions
ElevenLabs makes three performance claims. Each one has a footnote that limits what it means, so here they are together.
| Claim (per ElevenLabs) | What the footnote says it is based on |
|---|---|
| Ranked first by Artificial Analysis | The Artificial Analysis Provider Voice Arena leaderboard, September 2026 |
| Preferred by about 75% of listeners in blind head-to-head tests | ElevenLabs' blind preference tests in September 2026 against four named models: Cartesia Sonic 3.6, Inworld TTS-2, Google Gemini 3.8 Flash-Lite TTS and Google Gemini 3.8 Flash TTS. Graders heard the same line from both models and picked which was more expressive and which more natural; a tie counted as half |
| Eleven v4 Turbo: about 150 ms median time to first speech | ElevenLabs' September 2026 measurement of the time from request to audible speech, using identical scripts and default settings, against Cartesia Sonic 3.6, xAI TTS, Google Gemini 3.8 Flash-Lite TTS and OpenAI GPT-4o mini TTS. Network latency was measured and removed, and Turbo was tested over WebSocket streaming |
Two things are worth noticing:
- The 75% figure is a vendor-run test. It compares Eleven v4 with four models chosen by ElevenLabs. It says nothing about models that weren't in the test, and the post doesn't publish how many graders or lines were used.
- The post gives two latency figures. One section puts Eleven v4 Turbo's median inference latency at about 100 ms. The footnoted claim is about 150 ms median time to first speech, defined as request to audible speech. The post doesn't define "inference latency" separately, so treat them as two different numbers and say which one you are quoting.
For a recorded tutorial, latency barely matters. You render the voiceover once and play it back. The Turbo variant is aimed at live voice agents, not at pre-recorded narration.
What changed for tutorial narration
Here is what ElevenLabs says changed, mapped to the problems you hit when narrating software. The right-hand column paraphrases the announcement.
| Narration problem | What ElevenLabs says v4 changes |
|---|---|
| Every line sounds equally important | The model reads tone, pacing, emotion and context from the text itself |
| You can't tell the voice how to read a line | You can describe the delivery in plain language or with inline audio tags, and ElevenLabs says v4 follows them more closely than its earlier models |
| Product names and jargon come out wrong | Support for International Phonetic Alphabet (IPA) phonemes is improved, so custom pronunciations should behave more reliably |
| A regenerated line doesn't match the rest | Speaker identity holds more reliably across generations, narration and regenerated lines |
| Long scripts drift or break between chunks | Chaining generations together for long content ("request stitching") is more reliable |
| The localised version sounds like a different person | A voice recorded in one language can speak the others with a native-sounding accent while keeping its identity |
Directing delivery
The audio-tag examples in the announcement are mostly about character and sound effects, such as a laugh or a phone buzzing. ElevenLabs does mention narration among the uses where direction helps. For a tutorial, our reading is that the useful directions are much plainer:
- Slow down on the step that matters. A direction such as "read this calmly and a little slower" before the one action a viewer must not miss.
- Separate warnings from instructions. A slightly firmer delivery on "Make sure you save first" than on "Click Settings".
- Keep it neutral elsewhere. Over-directing a 40-step walkthrough makes it sound like an advert. Most lines should be plain.
The full tag syntax is in the ElevenLabs developer documentation, which the post links to. Check it before you write tags into a script, because tag behaviour is model-specific.
Pronunciation of product names
For software teams, this change is likely to matter most. Tutorials are full of words a general voice model has rarely heard: product names, feature names, acronyms and file formats. IPA support means you can specify how a word should sound instead of misspelling it on purpose to trick the voice.
ElevenLabs doesn't publish a measure of how much IPA handling improved, so test it on your own glossary rather than assume it.
Consistency across a series
A tutorial series is narrated over weeks, and individual lines get regenerated whenever the UI changes. If the voice shifts each time, the series sounds stitched together. ElevenLabs' claims about steadier speaker identity on regenerated lines, and more reliable stitching, are aimed at that. Only a test on your own voice and script length will show whether they hold.
What it means for localisation
If you publish tutorials in several languages, the claim that one voice can speak the other languages while keeping its identity is the headline. Microsoft made a similar claim when it announced MAI-Voice-2.1 on 1 October 2026, for 23 languages. That is two vendors making the same promise within days of each other. It doesn't tell you how widely the feature will spread, or what it will cost on any given plan.
For tutorials, a consistent voice across languages is only half the job. The other half is timing. A German sentence is often longer than its English source, so the translated narration may not finish before the screen moves on. That is a timing problem, not a voice problem, and a better voice model doesn't solve it.
Voice cloning: the consent question
Cloning from about 10 seconds of audio is convenient for narrating with your own voice. It also lowers the bar for cloning someone else's. The announcement doesn't describe consent checks for Instant Voice Clones. Before you clone a colleague's voice for a training series, get their written permission and agree how the voice may be used. Our guide to AI voice cloning and consent covers what to put in that agreement.
If you publish on YouTube, also read YouTube's own rules. Its disclosure help page lists cloning your own voice for voiceovers among the uses that don't need disclosure. We covered the details in YouTube's AI disclosure rules for tutorials.
How to test a new voice model on your tutorials
Benchmarks are run on someone else's sentences. Before you switch, run a short test on your own material.
- Pick three real scripts. A short one, a long one, and one full of product names.
- Build a pronunciation list. Ten to twenty terms your current voice gets wrong.
- Generate the same lines in both models. Same voice, same settings where possible.
- Regenerate one line in the middle. Listen to whether it still matches the lines around it.
- Listen against the video. Play the narration over the screen recording, not on its own. A voice that sounds great alone can feel too slow or too dramatic over a UI.
- Check one translated version. If you localise, test the language you publish most.
For a broader look at choosing a voice for screen content, see AI voiceover for videos and how to add AI voice to a screen recording.
Where Vorec fits
A better voice model improves how a line sounds. It doesn't write the script, and it doesn't make the narration line up with what happens on screen. Those are the parts that take most of the time when you narrate a tutorial.
Vorec records your screen. It has its own macOS recorder, and an AI agent can drive it for you through the Claude Code plugin, capturing locally for you to review before anything is uploaded. It then drafts narration matched to the workflow it captured and generates the voiceover, so nothing is spoken into a microphone. Freeze-sync can hold a frame to give an explanation time to finish, which helps when a line runs longer than the action it describes. The same capture can also produce a written step-by-step guide, and narration regenerates in supported languages on eligible plans without re-recording.
Vorec's Trial is $0 with 100 credits and is a 7-day free trial with no credit card required. Trial includes up to 3 projects; exports carry a watermark. Paid plans are Starter at $9 a month, Pro at $24 and Business at $59. Each paid plan has a launch allowance for its first three months (500, 1,500 or 5,000 credits a month), then a standard allowance (150, 500 or 1,500). Translation is not on the Trial; Starter covers 3 languages, and Pro and Business cover all supported languages. Prices checked 19 October 2026.
FAQ
When was Eleven v4 released?
ElevenLabs published its announcement on 28 September 2026.
What is the difference between Eleven v4 and Eleven v4 Turbo?
ElevenLabs describes Eleven v4 as its most emotive text-to-speech model, and Eleven v4 Turbo as the same technology tuned for low-latency uses such as voice agents. For pre-recorded tutorial narration, latency matters little.
How many languages does Eleven v4 support?
ElevenLabs says both models support more than 90 languages.
How much audio does Eleven v4 need to clone a voice?
ElevenLabs says an Instant Voice Clone can work from about 10 seconds of audio. Get the speaker's permission before cloning any voice that isn't yours.
Is Eleven v4 the best text-to-speech model?
ElevenLabs cites a first-place ranking on the Artificial Analysis Provider Voice Arena leaderboard in September 2026 and its own preference test against four named models. Those are its claims about those comparisons. Test it on your own scripts before deciding.
Can I use Eleven v4 for free?
The announcement invites readers to create a free account to try both models. It doesn't list per-plan limits for v4, so check ElevenLabs' pricing page for the current allowances.
Want the recording to narrate itself? Record with Vorec, or upload one you already have, and get a narrated tutorial plus a written guide. Start free.