ElevenLabs Scribe v3 Cuts Speech Errors 27%

ElevenLabs on Tuesday, Nov. 10, released Scribe v3 and Scribe v3 Live, two speech-to-text models that the voice AI company says make 27% and 43% fewer errors, respectively, than the versions they replace. The real-time model also labels who is speaking as it transcribes, a first for the Scribe line.

Key takeaways

• Scribe v3, for recorded audio, cuts word errors 27% versus Scribe v2 across 102 languages on ElevenLabs' internal version of the FLEURS benchmark.
• Scribe v3 Live, for real-time audio, lowers semantic word error rate 43% versus Scribe v2 Realtime and adds live speaker diarization.
• API list prices are unchanged from the previous generation: $0.20 an hour for Scribe v3 and $0.22 an hour for Scribe v3 Live.

How accurate is Scribe v3?

ElevenLabs measured Scribe v3 on an internal version of FLEURS, a set of short, clean read-speech clips in 102 languages. The company says word error rate (WER) fell 27% from Scribe v2 and improved in every language it tested.

Gains were 42% in German, 35% in Spanish and 23% in English. Harder languages also improved, with Swahili down 34%, Tamil 29% and Thai 28%.

For speaker labeling, Scribe v3 makes 12% fewer diarization errors than Scribe v2 overall and 17% fewer on recordings with three or more speakers.

What Scribe v3 Live changes for voice agents

The real-time model posted the larger jump. ElevenLabs says Scribe v3 Live reduced semantic WER, which counts an error only when meaning changes, by 43% and plain WER by 31% against Scribe v2 Realtime. Greek improved 77%, Danish 54%, Croatian 45% and Hindi 36%.

On the English portion of a 16-language dataset of real meetings, agent calls and voice notes that ElevenLabs built itself, Scribe v3 Live had a 3.1% diarization error rate. Other real-time models it tested scored 5.9% to 7.3%. The company says it will open-source the dataset "soon."

The model also predicts when a caller has finished speaking from conversational context, not just silence, so a caller who pauses halfway through reading out a number is not cut off. A multilingual mode handles speakers who switch languages mid-sentence.

Early customer test: Eneco

Dutch energy supplier Eneco, which serves more than 2 million homes and businesses, tested Scribe v3 Live on its voice agent before launch. In real customer calls, the agent recognized postcodes correctly on the first try 40% more often and house numbers 15% more often, according to ElevenLabs.

"It's great to see significant results this early, and we are excited to continue collaborating," said Ruud Huigsloot, a senior product manager at Eneco.

Price and developer controls

ElevenLabs' API pricing page lists Scribe v3 at $0.20 an hour and Scribe v3 Live at $0.22 an hour, the same rates as Scribe v2 and Scribe v2 Realtime. Keyterm prompting, which accepts up to 1,000 expected terms such as brand names or drug names, costs extra per hour. Scribe v3 Live adds a separately priced keyterm library that the documentation says can hold up to 100,000 terms.

Existing integrations switch by changing the model ID to scribe_v3 or scribe_v3_live.

Context: a crowded speech-to-text market

The release lands a week after rival Soniox shipped STT v6 for voice agents at a listed $0.12 an hour for real-time use. Big platforms are also pushing cheaper voice stacks, including Google's new voice models that talk while they think.

All of ElevenLabs' accuracy figures are its own benchmark results. The company illustrates the stakes this way: a one-hour meeting is about 9,000 words, so the gap between a 4% and a 2% error rate is 180 wrong words.

What's next

Both models run in the United States by default, with isolated EU, India and Singapore environments for enterprise customers. ElevenLabs says they are HIPAA-eligible with a business associate agreement, and it is also releasing a version of Scribe v3 Live tuned for medical conversations in ElevenAgents. Independent benchmarks and the promised open dataset have not yet been published.

FAQ

What is the difference between Scribe v3 and Scribe v3 Live? Scribe v3 transcribes recorded files, while Scribe v3 Live is a low-latency model for streaming audio, such as calls handled by voice agents.

How much does Scribe v3 cost? ElevenLabs lists $0.20 an hour for Scribe v3 and $0.22 an hour for Scribe v3 Live on its API pricing page, with add-on fees for keyterm features.

Are the accuracy claims independently verified? No. The 27% and 43% figures come from ElevenLabs' own tests on an internal FLEURS version and its own dataset.

References

• Introducing Scribe v3 and v3 Live, our most accurate speech to text models — ElevenLabs (Nov. 10, 2026)
• Models | ElevenLabs Documentation — ElevenLabs
• ElevenLabs API Pricing: Plans for Creators & Businesses of All Sizes — ElevenLabs