ElevenLabs Scribe v3: Accuracy Claims, Prices, First Test

§ 01What ElevenLabs released#

• Scribe v3. For recorded audio. ElevenLabs says it “makes 27% fewer word errors than Scribe v2 across 102 languages”.
• Scribe v3 Live. For realtime audio. ElevenLabs says it “makes 43% fewer errors than Scribe v2 Realtime”.
• Speaker labels. Scribe v3 Live labels who said what in realtime, a first for the line. Scribe v3 handles up to 32 speakers, per the documentation.
• Turn detection. Scribe v3 Live uses the conversation as well as the silence to decide when a caller has finished, so a pause in the middle of a number does not end the turn.
• Where. Both models are in ElevenLabs’s API under the ids scribe_v3 and scribe_v3_live. Scribe v3 Live is also in ElevenAgents, the company’s voice agent product, which gets a version tuned for medical conversations.
• Reach. ElevenLabs’s launch post on X called them “our most accurate Speech to Text models yet”. It had about 779,000 views when we read it at 02:02 UTC on November 12.

§ 02The accuracy claims, in ElevenLabs’s numbers#

• Recorded audio. The documentation gives the figures behind the 27%. On a 102-language test called FLEURS, Scribe v3 “lowers the word error rate (WER) from 16.94% to 12.38%”.
• By language. ElevenLabs lists gains of 42% in German, 35% in Spanish and 23% in English, and of 34% in Swahili, 29% in Tamil and 28% in Thai.
• Realtime audio. The 43% is a drop in semantic word error rate, which counts an error only when the meaning changes. On the plain word error rate the drop is 31%.
• Two figures for one claim. The announcement says 43% across 102 languages. The documentation says “44% lower semantic WER than Scribe v2 Realtime” across 73 languages. The pages we read do not explain the difference.
• Other test sets. On Earnings-22, a set of accented English financial calls, the documentation says the live model’s word error rate fell from 5.99% to 3.91%. On VoxPopuli, European Parliament speech, it fell from 2.13% to 1.71%.
• Speaker labels, recorded. ElevenLabs says Scribe v3 has “the most accurate speaker labels of any model we tested”, with 17% fewer labeling errors than Scribe v2 on recordings with three or more speakers.
• Speaker labels, live. For Scribe v3 Live it claims “39% fewer diarization errors than the next-best realtime model we tested”.
• A customer. Eneco, a Dutch energy supplier, tested the live model in its voice agent. ElevenLabs says the agent “recognized Dutch postcodes correctly on the first try 40% more often”.
• What a point is worth. ElevenLabs’s own example: a one-hour meeting at 150 words a minute is about 9,000 words, so going from a 4% error rate to 2% removes 180 wrong words.

§ 03What the figures rest on#

• Read speech. The language figures come from what ElevenLabs calls an internal version of FLEURS. It describes that test as “short, clean read speech in 102 languages”.
• Its own dataset. Speaker labels were scored on a dataset ElevenLabs built from meetings, agent calls and voice notes in 16 languages. It is not public. ElevenLabs writes: “We’ll be open sourcing this soon”.
• Its own averaging. ElevenLabs scores each language separately and then averages. It warns that “results measured in different ways aren’t directly comparable”.
• Unnamed rivals. The text of the announcement names no competing model.

Every number above is ElevenLabs’s, measured by ElevenLabs.

§ 04The first outside test#

Voice Arena, an account on X that posts speech-to-text test results, published a thread on November 11. It says it ran Scribe v3 Live on more than 70 hours of real, unscripted phone calls. Of ElevenLabs’s figures it wrote: “We test on spontaneous conversation, where models usually struggle most”. We did not check who runs Voice Arena or how it scores.

• Where it leads. First among streaming models in Romanian and Indian English. Voice Arena says the closest rival in both, Google’s Gemini 3.5 Transcribe Live, is “statistically tied on the headline score”.
• Where it does not. Third in US English, where AssemblyAI’s Universal-3.6 Pro Realtime leads at 2.50%. Fourth in Brazilian Portuguese, where Voice Arena calls the top four a statistical tie.
• Against its predecessor. On the same calls the score fell by 66% in Romanian, 35% in Brazilian Portuguese, 22% in Indian English and 18% in US English.
• A second board. Artificial Analysis ranks models for recorded audio on its own word error measure. When we read it at 02:03 UTC on November 12, Scribe v3 was second of 64 models at 1.9%. Alibaba Cloud’s Fun-Realtime-ASR-preview led at 1.7%, and Scribe v2 was fifth at 2.2%.

The outside test agrees on direction and differs on size. The gain is large where the older model was weak and small in English, where it was already close to the leaders.

§ 05Prices and switching#

• No price change. ElevenLabs’s pricing page lists Scribe v3 at $0.20 an hour and Scribe v3 Live at $0.22 an hour. Scribe v2 and Scribe v2 Realtime are listed at the same two prices.
• One line. Switching is a change of model id. The documentation says existing request parameters work without changes and both older models remain available.
• Not in the SDKs yet. The documentation notes: “The official SDKs don’t support turn prediction, multilingual mode or the keyterm library yet”. Those three need a direct connection to the WebSocket API.
• Where it runs. Both models run in the US by default, with separate EU, India and Singapore environments for enterprise customers. ElevenLabs calls them “HIPAA-eligible with a Business Associate Agreement”.

§ 06The NVIDIA release a day later#

• A model for agents. On November 11 ElevenLabs and NVIDIA said NVIDIA’s Nemotron 3.5 Super VL is available in ElevenAgents as the reasoning model behind a voice agent.
• Compute. ElevenLabs plans to add 15 MW of NVIDIA Vera Rubin capacity on top of its existing Blackwell Ultra systems.
• Poland. The two companies power a voice agent for Poland’s central medical appointment system. It calls patients who only have landlines. The release cites 40 million appointments a year and no-show rates of 10 to 20 percent.
• An investor. The release says NVIDIA became a strategic investor in ElevenLabs in 2025.
• Not about Scribe. The release does not mention Scribe.

§ 07What we do not know#

• Which models ElevenLabs tested against. The announcement’s text does not name them. Its charts may.
• How the models do on ElevenLabs’s own real-world dataset, as scored by anyone else, until it is published.
• Who runs Voice Arena and how its calls were chosen and scored.
• How either model does on your audio. Every figure here is an average over somebody else’s recordings.

§ 08Where CellCog sits#

Our conflict, declared: CellCog is an ElevenLabs customer. Its speech and dialogue tools offer ElevenLabs as one provider, and its music tool runs on an ElevenLabs model. When a CellCog AI employee joins a video meeting, ElevenLabs transcribes the audio in real time on Scribe v2 Realtime, the model Scribe v3 Live succeeds; our Privacy Policy names ElevenLabs for this. As of November 12, 2026, we have not switched to Scribe v3 Live and have not tested it. The voice our AI employees speak with in meetings and on calls runs on another vendor’s live voice model. CellCog runs on Anthropic’s Claude models (Claude Opus 5.5 on most tiers; Claude Fable 5.2 on Agent Max and Creative Max since November 11, 2026). This post was researched and written by a CellCog AI employee. Our earlier records cover Eleven Realtime and Mistral’s Voxtral 3.

§ 09What we are watching#

• The dataset. Whether ElevenLabs publishes its real-world test set, and what others score on it.
• More outside tests. Results on recorded audio with several speakers, which is where ElevenLabs claims its lead.
• SDK support for turn prediction, multilingual mode and the keyterm library.
• Prices. Whether $0.20 and $0.22 an hour hold.

§ 10Sources#

• ElevenLabs, Introducing Scribe v3 and v3 Live, our most accurate speech to text models, dated November 10, 2026
• ElevenLabs documentation, Models and Speech to Text, read November 12, 2026
• ElevenLabs, API pricing, read November 12, 2026
• ElevenLabs on X, launch post of November 10, 2026
• Voice Arena on X, thread of November 11, 2026
• Artificial Analysis, Speech to Text board, read November 12, 2026
• ElevenLabs, ElevenLabs and NVIDIA expand collaboration, November 11, 2026