Scribe v3 cuts transcription errors 27%. Your brand still needs a list

Pull up the automatic notes from your last meeting. I'd bet your product name is misspelled somewhere. Or your biggest client's name. Or “fifty” where someone said “fifteen”. Sounds trivial. It isn't, the moment an AI turns those notes into tasks, CRM entries or a summary for leadership.

What's new

• Scribe v3: On November 10, 2026, ElevenLabs launched a new model for recorded audio. According to the company it makes 27% fewer word errors than Scribe v2 across 102 languages, with German improving by 42% and English by 23%.
• Scribe v3 Live: the realtime version, with 43% lower semantic word error rate than Scribe v2 Realtime and speaker diarization for the first time, plus a multilingual mode for conversations that switch languages mid-sentence.
• Who said what: 12% fewer diarization errors than v2, and 17% fewer with three or more speakers.
• Keyterm lists: keyterm prompting takes up to 1,000 terms such as brand names, acronyms and jargon, plus formatting rules written in plain English.
• Access: both models via the ElevenLabs API, v3 Live also in ElevenAgents. The post doesn't list prices.

All numbers come from ElevenLabs' own testing, partly on an internal version of the FLEURS benchmark, which uses short, clean read speech. A real team meeting is harder.

The transcript is the foundation

My rule of thumb: AI doesn't fix broken ways of working, it widens the cracks in the foundation. With meeting AI, the foundation is the transcript. Everything downstream, the summary, the to-dos, the agent drafting the follow-up email, inherits every mishearing. ElevenLabs does the math itself: a one-hour meeting is about 9,000 words, so the gap between a 4% and a 2% error rate is 180 wrong words. And errors cluster in names, numbers, acronyms and jargon, the company says. Exactly the words that matter.

Your keyterm list is a brand asset

Here's the second-order point for me. An agent only knows what you give it. That applies to listening, too. Those 1,000 terms are really your glossary: product and brand names, client names, internal acronyms, the industry slang every marketing meeting runs on. Maintain that list and you get cleaner data. Skip it and you get a faster model that keeps getting your name wrong.

The catch

Scribe is a building block for developers, not a finished meeting tool. Whether your note-taker runs on ElevenLabs, and whether it lets you add your own vocabulary, is a question for your vendor.

What I'd do now

• → Take three real meeting transcripts and count how often brand, product and client names are wrong. That's your baseline.
• → Build a list of your 100 to 200 most important names and acronyms. Maintain it once, reuse it in every tool.
• → Ask your meeting-tool vendor: is there custom vocabulary, and which speech model runs underneath?

Bottom line: Even the best speech model only knows your brand name once you write it down.