Transcribing audio used to be the chore you kept putting off until Friday afternoon. ElevenLabs' new speech recognition model, Scribe v3, is built to make that chore disappear: 102 languages, 27% fewer word errors than the previous generation, and a live version that can switch languages mid-sentence. It's out now, and you can reach it through the website, the API, or the company's agent platform.

What happened

ElevenLabs released Scribe v3, the next version of its speech-to-text engine, and split it into two products that solve two different problems.

One engine, two modes

Scribe v3 handles recordings you already have — a Zoom call, a voice memo, a podcast episode, an interview, a lecture. You hand it a file, it hands you text.

Scribe v3 Live listens while people are still speaking. That's the version you want for a live call, a webinar, or any app where captions have to show up in the moment. Live is built into ElevenAgents, ElevenLabs' platform for voice agents.

Both models are available through ElevenAPI, so anyone building a product can plug them in directly.

The three numbers worth remembering

102 languages. Not a short demo list with 'more coming soon' tacked on the end — support for 102 languages.

27% fewer word errors. The measure is WER, or word error rate. Picture a dictation test: someone reads a paragraph out loud and a teacher counts every word you got wrong. That share of wrong words is your WER, and lower is always better. A 27% cut is a real step, not a marketing round-down.

43% fewer meaning errors. This one is the semantic WER, and it's reported for the Live model. It's the gap between getting every word right and getting the point right. 'Ship it Tuesday' and 'ship it Thursday' are one word apart and a week apart in real life. Semantic scoring catches misses like that, which is exactly what ruins meeting notes and call summaries.

Two smaller upgrades that save real time

Live can now switch languages on the fly, so a call that drifts from English into Spanish and back doesn't collapse halfway through. And you can set formatting rules ahead of time — tell it 'write dates as YYYY-MM-DD' and it will, every time, instead of handing you three different date styles in one transcript.

What it means for you

Accuracy percentages are abstract until you map them onto your own week. Here's how it plays out.

At home

Old family videos, voice notes from your grandmother, a recipe someone dictated while cooking — all of it turns into searchable text. A 27% drop in word errors is the difference between a transcript you can skim and one you have to re-listen to three times.

At work

Meetings stop being a memory test. You record, you get text you can search for 'who agreed to what,' and the meaning-level gain matters more than the word-level one: the notes say what was decided, not just what was said out loud.

For your business

Support calls, sales calls, customer interviews, onboarding sessions — one engine covering 102 languages means you don't need a different tool for every market you sell into. Once you've got clean text, free AI tools like the ones at MyKreaTool can turn a wall of transcript into a summary, a blog post, or a set of FAQs without you typing a word of it.

For studying

Lectures, seminars, group projects, language practice. Live mode with on-the-fly switching is genuinely handy if you're learning a language: you hear the sentence and read it at roughly the same time, which is how the brain likes it.

For creative work

Podcasts, video subtitles, interviews for a book, first drafts dictated while you walk the dog. The formatting rules are the quiet hero here — if your publisher wants dates as YYYY-MM-DD, you set it once and stop fixing it by hand.

For extra income

Freelance transcription and subtitling, course transcripts, YouTube localization into other languages. Accuracy is the part clients actually pay for, because cleanup is the part that eats the hours.

How to try it right now

Start on the website before you go anywhere near an API key.

1. Open ElevenLabs Scribe v3 and sign in. No install, no code, no credit card dance just to see what it does.

2. Upload a short clip first — sixty seconds is plenty. Pick something messy: two people talking over each other, background noise, a name that's hard to spell.

3. Check the language. With 102 supported, auto-detect usually gets it, but setting it manually is faster when you already know.

4. Set one formatting rule. Try 'dates as YYYY-MM-DD' or 'numbers as digits.' This is the feature people skip and then wonder why their transcripts look like a ransom note.

5. Read the output against the audio. Listen for the words that matter to you — names, figures, deadlines. That's your real accuracy test, not a benchmark table.

6. Scale up once you're satisfied. Longer files for Scribe v3, or route live audio through ElevenAgents if you need captions in real time.

7. Builders: the ElevenLabs speech-to-text documentation covers the ElevenAPI endpoints, parameters, and formatting options.

Upsides and what changes

The headline win is coverage. One model, 102 languages, one integration — that kills a whole category of 'we'll need a separate vendor for Poland' conversations.

The second win is the 43% semantic figure. Word-level accuracy is table stakes by now; meaning-level accuracy is what decides whether a summary is trustworthy enough to send to a client without reading the transcript yourself first.

Third, the split between recorded and live is the right shape for the product. Most people need one or the other, and pretending a single mode does both well is how tools end up mediocre at both.

Fourth, preset formatting rules are a small feature with outsized value. Anyone who's spent an afternoon find-and-replacing date formats knows exactly what that's worth.

Limitations

A few honest caveats. The 27% and 43% numbers come from ElevenLabs comparing its new models against its older ones, so read them as 'better than the last version' rather than 'best on the market' — independent benchmarks will tell that story later. Supporting 102 languages doesn't mean all 102 are equally polished; long-tail languages nearly always lag the big ones, so test yours with real audio before you commit. Live transcription is also harder than file transcription by nature, because it can't rewind and re-listen to a sentence the way a file-based model can, so expect more cleanup on the live output. And the announcement led with accuracy and language coverage, not pricing, so check current plans on the site before you build a workflow that depends on it.

Conclusion

Speech recognition has quietly become the plumbing behind a lot of work — meetings, subtitles, notes, support tickets, localization — and Scribe v3 pushes on the two things that actually bother people: languages you can't use and errors you have to fix by hand.

One thing to do today: grab a one-minute recording of your own voice or your last meeting, run it through Scribe v3 on the ElevenLabs site, and read the output against the audio. Ten minutes of testing will tell you more than any benchmark table.