Scribe v3 Cuts Recorded Transcription Errors 27% Across 102 Languages

ElevenLabs positions Scribe v3 for recorded audio and Scribe v3 Live for real-time transcription. Their headline accuracy claims are separate comparisons: Scribe v3 makes 27% fewer word errors than Scribe v2 across 102 languages, while Scribe v3 Live makes 43% fewer than Scribe v2 Realtime.

For recorded audio, ElevenLabs also reports up to a 17% improvement in diarization on recordings with three speakers, and the lowest diarization error among the models it tested. Diarization identifies who said what. That matters when a transcript must preserve not just the words, but which speaker contributed them.

Scribe v3 also gives users control over transcript formatting. They can enter rules in plain English—for example, formatting dates as YYYY-MM-DD—and choose whether numbers appear as digits or words. A keyterm library can hold up to 10,000 names or technical terms. These controls address different sources of transcription friction: inconsistent formatting, unfamiliar vocabulary and domain-specific language.

Scribe v3 Live is designed for conversations as they happen. It adds speaker labels to live audio and uses the context of the conversation to detect turns, rather than relying only on silence. ElevenLabs says that distinction can prevent a caller who pauses halfway through a phone number from having the turn cut off.

In a phone-call demonstration, the transcript distinguishes an agent from a client and renders the caller’s policy identifier as “HM-4470219,” although the spoken version spells out its letters and digits. The example illustrates the practical role of speaker labeling and number recognition in a live exchange; it does not establish performance beyond the company’s reported evaluation.

ElevenLabs says Scribe v3 Live had fewer errors than every other real-time model it tested. The launch chart ranks it ahead of Grok Voice, Gemini 3.5 Transcribe Live, GPT-Live-Transcribe and Deepgram Nova-3. Those comparisons are the company’s test results; the source does not provide further evaluation details.

For conversations that switch languages mid-sentence, ElevenLabs offers a multilingual mode. Its demonstration moves between English and Korean within turns, including a discussion of whether to launch a feature the following Monday or delay it until Wednesday. The relevant capability is continuity across a language switch, rather than translating or summarizing the exchange.

The product’s formatting and keyterm controls address another challenge: preserving details whose exact form matters. A technical-note example includes a gridline, competing rebar measurements and a scheduled pour time. The transcript renders those details as “C-7,” “150 mm,” “125” and “6:30.” The example shows the kind of information the product is intended to capture; ElevenLabs describes Scribe v3 as paying close attention to words and numbers.

The clearest customer outcome in the supplied launch material comes from Eneco, which supplies energy to more than 2 million homes and businesses. Eneco tested Scribe v3 Live on its voice agent across more than 1,000 customer calls. The company reports that the agent recognized Dutch postcodes correctly on the first try 40% more often. This result is specifically tied to the live model’s use in a voice-agent deployment.

Scribe v3 Live is available in ElevenAgents, while both models are available in ElevenAPI. Scribe v3 is now the default in the ElevenLabs app. The deployment options mirror the product distinction: a model for recorded audio and a separate one for real-time use.