ElevenLabs' Scribe v3 Cuts Live Transcription Errors by 43% Across 102 Languages

ElevenLabs has released Scribe v3 for recorded audio and Scribe v3 Live for streaming transcription, according to its launch post. The models target overlapping speakers, mid-sentence pauses, specialized vocabulary, and code-switching between languages. Scribe v3 Live also adds speaker diarization, which assigns each speech segment to a speaker.

Migration takes one model ID

ElevenLabs lists both models in ElevenAPI and Scribe v3 Live in ElevenAgents. Existing REST integrations can set model_id to scribe_v3 for files or scribe_v3_live for streams. The JavaScript SDK uses camel-case property names such as modelId and commitStrategy.

import { ElevenLabsClient } from "@elevenlabs/elevenlabs-js";

const elevenlabs = new ElevenLabsClient();

const connection = await elevenlabs.speechToText.realtime.connect({
  modelId: "scribe_v3_live",
  commitStrategy: "turn_prediction",
});

A model-ID change preserves the existing API shape, although transcripts may differ in segmentation, formatting, speaker labels, and turn timing. Scribe v2 and Scribe v2 Realtime remain available, allowing teams to compare outputs before migrating production traffic.

The published comparisons quantify accuracy rather than end-to-end latency or cost. Evaluations of Scribe v3 Live will still need workload-specific measurements for median latency, tail latency, and current per-minute pricing.

Where the errors fell

ElevenLabs measured multilingual accuracy with an internal version of FLEURS, a benchmark built from clean, read speech in 102 languages. A separate dataset covers meetings, agent calls, and voice notes in 16 languages for diarization testing; the company says it plans to open-source that dataset. These internal results have not been independently verified.

Word error rate, or WER, counts substitutions, deletions, and insertions relative to a reference transcript. ElevenLabs also reports a semantic WER that discounts mistakes when the intended meaning remains intact. Diarization error rate, or DER, measures mistakes in speaker assignment. Lower values indicate fewer errors.

On recordings with at least three speakers, Scribe v3 produced 17% fewer diarization errors than Scribe v2. For English streaming audio, Scribe v3 Live recorded a 3.1% DER, compared with a range of 5.9% to 7.3% for the other real-time models in the test.

The averaging method affects comparisons with published results from other vendors. ElevenLabs scored each language separately and then averaged all 102 scores, giving every language equal weight. Pooled evaluations let languages with more clips contribute more heavily and often produce lower aggregate error rates.

Because FLEURS contains clean, read speech, its WER results do not establish performance on noisy calls or heavily overlapping conversations. The separate real-world dataset evaluates diarization, leaving production audio tests necessary for domain-specific accuracy.

Controls for pauses, jargon, and code-switching

Scribe v3 Live adds three controls for common voice-agent failure modes:

• commit_strategy=turn_prediction uses conversational context to decide when an utterance has ended, reducing premature commits when a caller pauses during a number or sentence.
• keyterm_library stores thousands of product names, medical terms, and other domain-specific phrases, consulting them when the audio suggests a likely match.
• multilang_mode=true supports language changes within a single sentence.

For recorded audio, Scribe v3 adds numbers_format for choosing digits or written numbers and secondary_language_codes for multilingual recordings. Batch requests also support prompts containing up to 1,000 known terms and plain-language transcript rules such as formatting every date as YYYY-MM-DD.

Two percentage points add up

At 150 words per minute, a one-hour meeting contains about 9,000 words. A 4% WER corresponds to roughly 360 edit operations, while a 2% WER corresponds to about 180. That difference averages three fewer errors per minute, with names, account numbers, acronyms, and technical terms carrying disproportionate operational risk.

Eneco’s customer-call case study reports that an agent using the new model recognized Dutch postcodes correctly on the first attempt 40% more often and house numbers 15% more often. A misheard digit can block the identity check at the start of a call, although the published case study does not provide the absolute success rates behind those relative gains.

Regional processing and retention controls

ElevenLabs says both models process data in the United States by default. Enterprise customers can use isolated environments in the European Union, India, and Singapore to meet data-residency requirements. Zero Retention Mode deletes audio and transcript data after a request completes.

The company lists the service as SOC 2, ISO 27001, and PCI DSS Level 1 compliant. It is also HIPAA-eligible when covered by a business associate agreement. A medical-optimized version of Scribe v3 Live is available through ElevenAgents documentation for clinical conversations.