ElevenLabs Launches Scribe v3 and Scribe v3 Live Speech-to-Text Models: 27% Fewer Word Errors Than v2 Across 102 Languages, First Realtime Speaker Diarization, and API Pricing of $0.20/$0.22 per Hour

Background and Industry Pain Points

Meeting notes, contact-center QA, voice agents and clinical documentation all start from a transcript, and every wrong word travels downstream. Real audio has background noise, people talking over each other, mid-sentence pauses and code-switching, which is exactly where most speech-to-text models fail. ElevenLabs says Scribe has already processed hundreds of millions of hours, and v3 targets these pain points.

Architecture Highlights and Mechanisms

ElevenLabs has not disclosed the architecture or parameter count. Functionally there are three changes. First, v3 Live brings speaker diarization to realtime streams for the first time. Second, turn detection uses conversational context rather than silence alone, the same system ElevenAgents uses. Third, there is more control: v3 supports language hints, numbers as digits or words, keyterm prompting with up to 1,000 terms, and formatting rules written in plain English, while v3 Live adds a multilingual mode for mid-sentence language switches and a keyterm library that holds thousands of terms.

Benchmarks and Measured Results

All of the following are vendor figures. On 102 FLEURS languages, v3 lowers WER by 27% versus v2 and v3 Live lowers semantic WER by 43% versus v2 Realtime; ElevenLabs says both beat every comparable model it tested, without listing rival scores. Diarization was measured on an in-house 16-language dataset of real meetings. Third-party Voice Arena tested v3 Live on 70+ hours of spontaneous phone calls and ranked it first in Romanian and Indian English. Note that ElevenLabs scores each language separately before averaging, so its numbers are not directly comparable with pooled leaderboards.

Developer Quick Start

In ElevenAPI, switch the model ID to scribe_v3 or scribe_v3_live. The batch API adds numbers_format and secondary_language_codes parameters, supported only by v3. At $0.20 / $0.22 per audio hour, 1,000 hours of audio costs about $200. Singapore teams can request local data residency and enable Zero Retention Mode. Benchmark on your own real call samples against v2 before migrating.