ElevenLabs Launches Scribe v3 and Scribe v3 Live with 27% Lower Error Rates and Realtime Speaker Diarization
ElevenLabs has officially released Scribe v3, cutting word error rates by 27% across 102 languages, alongside Scribe v3 Live featuring realtime speaker diarizat
On November 11, 2026, ElevenLabs officially launched two new Speech-to-Text models: Scribe v3 and Scribe v3 Live. Built for recorded audio, the batch model Scribe v3 reduces word error rate (WER) by 27% compared with Scribe v2 across 102 languages, while the low-latency streaming model Scribe v3 Live cuts semantic WER by 43% against Scribe v2 Realtime. Both models are now the default in the ElevenLabs app and are available immediately in ElevenAPI and ElevenAgents.

Image source: https://elevenlabs.io/blog/scribe-v3
For audio-first products such as meeting note-takers, clinical documentation tools, and contact-center voice agents, getting the initial transcript right is critical because every recognition error propagates downstream where it is far costlier to fix. ElevenLabs stated that the new models target messy real-world conditions—such as overlapping speakers in a crowded meeting, callers pausing mid-sentence on a noisy street, and conversations where speakers switch languages.
Multilingual Accuracy Gains and Batch vs. Realtime Evaluation
ElevenLabs evaluated both models against Scribe v2, Scribe v2 Realtime, and competing Speech-to-Text models across identical test sets. Multilingual transcription accuracy across 102 languages was measured on an internal version of the FLEURS benchmark, scoring each language individually before averaging so all 102 languages carry equal weight.
- Scribe v3 (Batch model for recorded audio): Reduces WER by 27% over Scribe v2 across 102 languages, with gains in every tested language. Error rates fell by 42% in German, 35% in Spanish, and 23% in English, alongside notable improvements in harder-to-transcribe languages such as Swahili (34%), Tamil (29%), and Thai (28%). On recordings with three or more speakers, Scribe v3 reduces speaker diarization errors by 17% compared with Scribe v2 (and by 12% overall).
- Scribe v3 Live (Realtime streaming model): Because streaming models emit text before hearing an entire sentence, ElevenLabs leads with semantic WER—which counts errors only when meaning changes rather than penalizing formatting or punctuation differences. Across 102 languages, Scribe v3 Live reduces semantic WER by 43% and standard WER by 31% compared with Scribe v2 Realtime, with improvements of 77% in Greek, 54% in Danish, 45% in Croatian, and 36% in Hindi.
Two methodological caveats apply to these benchmarks. First, because FLEURS contains single-speaker read speech, ElevenLabs evaluated diarization on a separate internal dataset of real meetings, agent calls, and voice notes across 16 languages. Second, to isolate speaker assignment accuracy from missed words in streaming transcription, realtime diarization scores were computed after aligning each model's transcript to the reference text; on the English portion of that dataset, Scribe v3 Live achieved a 3.1% diarization error rate versus 5.9% to 7.3% for other tested realtime models, though this aligned metric cannot be directly compared with batch diarization scores.
Realtime Speaker Diarization and Context-Aware Turn Detection
Scribe v3 Live introduces speaker diarization to ElevenLabs' realtime transcription lineup for the first time, labeling who said what on the fly with 39% fewer diarization errors than the next-best realtime model tested by the company.
The model also replaces traditional silence-timer turn-taking with context-aware turn detection. Conventional voice agents rely on fixed silence thresholds, forcing an awkward trade-off between cutting callers off when they pause mid-sentence to read a number and making the agent feel sluggish. By combining conversational context with audio cues, Scribe v3 Live predicts when a speaker has actually finished talking—a capability previously used in ElevenAgents and now exposed directly through ElevenAPI.
Both models expand output customization with Keyterm prompting (supporting up to 1,000 brand names, acronyms, or domain terms), configurable digit-versus-word number formatting, and plain-English formatting rules (such as enforcing YYYY-MM-DD dates). Scribe v3 Live additionally offers a multilingual mode for mid-sentence code-switching and a keyterm library capable of holding thousands of vocabulary items that the model consults only when it hears a likely match.
Enterprise Deployment and Real-World Voice Agent Results
Ahead of general availability, Dutch energy provider Eneco—which supplies energy to more than 2 million homes and businesses in the Netherlands—tested Scribe v3 Live on its ElevenAgents voice agent. In live customer calls, the agent recognized Dutch postcodes correctly on the first try 40% more often and house numbers 15% more often.
Taken together, context-aware turn detection and realtime diarization address a common bottleneck in automated phone support: when callers pause while dictating multi-digit postcodes or speak amid background chatter, holding the turn open while isolating the caller's voice prevents premature cutoffs during initial verification.
Existing ElevenAPI integrations can upgrade via a one-line change to the model ID (scribe_v3 or scribe_v3_live), and ElevenLabs is also releasing a version of Scribe v3 Live tuned for medical conversations inside ElevenAgents.
import { ElevenLabsClient } from "@elevenlabs/elevenlabs-js";
const elevenlabs = new ElevenLabsClient();
const connection = await elevenlabs.speechToText.realtime.connect({
modelId: "scribe_v3_live",
commitStrategy: "turn_prediction",
});
For enterprise compliance and data residency, both models run in the US by default with isolated deployment options in the EU, India, and Singapore, alongside a Zero Retention Mode that deletes audio and transcripts immediately upon request completion. They are covered under ElevenLabs' SOC 2, ISO 27001, and PCI DSS Level 1 programs, are HIPAA-eligible with a Business Associate Agreement, and support private deployment.
Sources
- ElevenLabs Official Blog: Introducing Scribe v3 and Scribe v3 Live