Published: 11 November 2026
ElevenLabs released Scribe v3 and Scribe v3 Live on 10 November, new speech-to-text models for recordings and live conversations. The company says Scribe v3 makes 27% fewer word errors than Scribe v2 across 102 languages, while the API price per hour of audio stays the same. Both models are now the default in the ElevenLabs app.
Scribe v3 Live now tells speakers apart during a live call
Scribe v3 is built for recorded audio such as meetings and voice notes. Scribe v3 Live works in real time, for example on a phone line, and also runs inside ElevenAgents, the company’s voice agent platform.
The main new feature is live speaker labels, known as diarization. This is the first time the Live version of Scribe offers them. ElevenLabs also says the model now judges better when a caller has finished speaking, because it weighs the rest of the conversation as well as the audio. If that holds up, a voice agent should interrupt callers less often. A multilingual mode follows speakers who switch language in the middle of a sentence.
The 27% comes from ElevenLabs’ own version of FLEURS
ElevenLabs reports 27% fewer word errors for Scribe v3 than for Scribe v2, averaged over 102 languages. For Scribe v3 Live it reports 43% fewer errors than Scribe v2 Realtime. Both figures come from an internal version of FLEURS, a public test set of short, clean sentences read aloud. ElevenLabs says it will soon open source a second test set of real meetings, agent calls and voice notes in 16 languages. Until then, nobody outside the company can check the numbers.
ElevenLabs also says Scribe v3 makes fewer word errors than every other batch model it tested, but it does not name those models. The examples it gives include 42% fewer errors in German and 23% fewer in English.
Eneco’s voice agent got Dutch postcodes right first time 40% more often
The most practical figure comes from Eneco, a Dutch energy supplier with more than 2 million homes and businesses as customers. Eneco runs its voice agent on ElevenAgents and tested Scribe v3 Live on real customer calls before launch. The agent recognised postcodes correctly on the first try 40% more often, and house numbers 15% more often. ElevenLabs does not say how many calls were tested or which model served as the baseline.
Details like these decide whether a call succeeds. A voice agent that mishears a postcode has to ask again, or it looks up the wrong address.
How much does Scribe v3 cost?
Scribe v3 costs $0.20 per hour of audio in the API and Scribe v3 Live costs $0.22. That is the same as Scribe v2 and Scribe v2 Realtime, according to the ElevenLabs API pricing page. The announcement itself mentions no prices.
- Model IDs: scribe_v3 for recordings and scribe_v3_live for real time. Moving from Scribe v2 means changing the model ID.
- Custom vocabulary: keyterm prompting takes up to 1,000 terms, such as product names, for an extra $0.05 per hour on Scribe v3 and $0.08 on Scribe v3 Live.
- Formatting: you choose digits or words for numbers, and you can set rules in plain English, such as one format for every date.
Audio stays in the EU only on an Enterprise contract
Scribe v3 and Scribe v3 Live run in the US by default. ElevenLabs offers isolated environments in the EU, India and Singapore, but only to Enterprise customers. The EU environment needs its own account, API key and endpoint. Its documentation says storage then stays in the EU, while processing stays in the EU only when you also use Zero Retention Mode through the API. That mode is available to select Enterprise customers. Transcript editing does not work in the isolated environments yet.
For European companies that run customer calls through a voice agent, this is a GDPR question. A postcode with a house number is an address, and so personal data.
For organisations building voice agents for customer service, the risk rarely sits in ordinary sentences. It sits in the details the process depends on: postcodes, customer numbers, product names and people’s names. An average error rate across 102 languages says little about those, which is why Eneco’s result is more useful than the headline figure.
If you already use Scribe v2, switching costs one changed line of code and nothing extra per hour. Before you switch, run your own recorded calls through both models. Count how often the details your process needs come out right first time, with and without keyterm prompting. If your callers are in the EU and you have no Enterprise contract, also compare Scribe v3 Live with an EU-hosted option such as Mistral’s Voxtral 3.
For more information, visit the official announcement of Scribe v3 and Scribe v3 Live on the ElevenLabs blog.
