---
type: "post"
id: "elevenlabs-ships-scribe-v3-and-scribe-v3-live-6339e7"
slug: "elevenlabs-ships-scribe-v3-and-scribe-v3-live-6339e7"
title: "ElevenLabs ships Scribe v3 and Scribe v3 Live speech-to-text models"
description: "ElevenLabs says Scribe v3 and Scribe v3 Live are now available in its Speech to Text API, for recorded audio and realtime transcription respectively."
retrieval_nugget: "ElevenLabs says Scribe v3 and Scribe v3 Live are now available in its Speech to Text API, for recorded audio and realtime transcription respectively. The release reports lower FLEURS error rates than their predecessors and adds controls for mixed-language speech, number formatting, turn-aware commits and domain-specific terms."
published_at: "2026-11-16"
updated_at: "2026-11-16"
record_date: "2026-11-16"
date_kind: "published_at"
topics: ["models"]
entities: ["ElevenLabs","Scribe v3","Scribe v3 Live","FLEURS"]
related_patterns: ["generative-media-becomes-application-infrastructure"]
novelty_label: "new_fact"
what_changed: "ElevenLabs has shipped new Scribe v3 and Scribe v3 Live speech-to-text models."
prior_refs: []
source_url: "https://elevenlabs.io/docs/changelog/2026/11/9"
source_urls: ["https://elevenlabs.io/docs/changelog/2026/11/9"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:elevenlabs-ships-scribe-v3-and-scribe-v3-live-6339e7"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/elevenlabs-ships-scribe-v3-and-scribe-v3-live-6339e7/","markdown":"https://newruntime.com/posts/elevenlabs-ships-scribe-v3-and-scribe-v3-live-6339e7.md","json":"https://newruntime.com/posts/elevenlabs-ships-scribe-v3-and-scribe-v3-live-6339e7.json"}
markdown_url: "https://newruntime.com/posts/elevenlabs-ships-scribe-v3-and-scribe-v3-live-6339e7.md"
---

# ElevenLabs ships Scribe v3 and Scribe v3 Live speech-to-text models

## Retrieval answer

ElevenLabs says Scribe v3 and Scribe v3 Live are now available in its Speech to Text API, for recorded audio and realtime transcription respectively. The release reports lower FLEURS error rates than their predecessors and adds controls for mixed-language speech, number formatting, turn-aware commits and domain-specific terms.

## Summary

ElevenLabs says Scribe v3 and Scribe v3 Live are now available in its Speech to Text API, for recorded audio and realtime transcription respectively. The release reports lower FLEURS error rates than their predecessors and adds controls for mixed-language speech, number formatting, turn-aware commits and domain-specific terms.

## What happened

ElevenLabs’ November 9, 2026 changelog announces two speech-recognition models: Scribe v3, with model ID `scribe_v3`, and Scribe v3 Live, with model ID `scribe_v3_live`. Both are available in the company’s Speech to Text API. The changelog positions Scribe v3 for recorded audio and Scribe v3 Live for realtime transcription. Scribe v2 and Scribe v2 Realtime remain available, so the announcement describes successors rather than an immediate removal of the older models.

For Scribe v3, ElevenLabs reports 27% fewer word errors than Scribe v2 on the FLEURS benchmark across 102 languages. The release adds a `numbers_format` parameter that can render numbers as digits or words. It also adds `secondary_language_codes`, for audio that mixes languages. These options address transcription output and language identification for a batch recording without requiring a change to the basic recorded-audio workflow.

For Scribe v3 Live, the company reports a 44% lower semantic word error rate than Scribe v2 Realtime on FLEURS across 73 languages. Its new `commit_strategy=turn_prediction` option commits a segment when the speaker has finished a turn rather than at the first pause. The changelog presents this as a way to avoid treating an ordinary pause as the end of a speaker’s contribution. A multilingual mode writes each language in its own script, while the keyterm library supports up to 10,000 terms.

The API changes are specific to the two paths. The create-transcript endpoint accepts `scribe_v3` as its model ID and the new number-format and secondary-language parameters. The realtime WebSocket accepts `scribe_v3_live`, along with turn prediction, multilingual mode and the keyterm library. The changelog does not state a separate price or a new access tier for either model; it describes them as available in the Speech to Text API.

## What the archive adds

New Runtime’s October 22, 2026 [earlier coverage of Eleven Realtime](https://newruntime.com/posts/introducing-eleven-realtime-a-simpler-way-to-build-8e4a6b/) said voice-agent model releases are becoming easier to evaluate by workload, while benchmark scores alone do not settle deployment fit. This is a new event in the same company’s speech and voice product line, not a restatement: Scribe v3 concerns speech recognition, not the previously covered voice-agent model.

## Trend signal

This release supports a workload-specific approach to generative-media infrastructure: recorded transcription and live transcription have distinct model paths and controls. New Runtime’s [generative-media pattern](https://newruntime.com/patterns/generative-media-becomes-application-infrastructure/) concerns media as composable application infrastructure, but this announcement does not establish that realtime voice models replace chained speech-to-text and text-to-speech systems. The pattern is relevant as a direction, not proof of that stronger claim.

## Why now

ElevenLabs dates the changelog entry November 9, 2026, and makes both models available through its Speech to Text API. That places the benchmark claims alongside an immediate product change for teams already using the API, rather than presenting the results only as research.

## Why it matters

Teams can now evaluate separate choices for recorded and realtime transcription rather than assuming one recognizer fits both latency profiles. For multilingual recordings, the secondary-language parameter and per-language scripts could make transcripts more usable downstream; keyterms may help preserve names and specialized vocabulary. Turn prediction is particularly relevant to live applications where pauses occur inside a speaker’s turn: waiting for a predicted turn boundary may produce more coherent committed segments than committing at the first silence.

This separation matters for system design. A recorded-meeting pipeline can optimize for a complete transcript and select explicit secondary languages before processing begins. A voice application, by contrast, must decide when text is stable enough to hand to downstream logic, display to a user, or use as input to another model. The `turn_prediction` setting gives teams a product-level boundary to test instead of treating every short silence as an instruction to act.

The benchmark percentages are a starting point, not a deployment decision. A team should test its own audio, language mix, vocabulary and latency requirements, and compare the resulting transcript quality against its current recognizer. The announcement makes the API capabilities more concrete, but it does not establish how they perform on a particular customer’s calls, noisy environments or domain-specific recordings.

## Limits and uncertainty

The accuracy improvements are ElevenLabs’ reported comparisons on FLEURS; the changelog does not give detailed benchmark methodology, per-language results, confidence intervals or independent replication. It also does not quantify latency, streaming stability or the effect of turn prediction on end-to-end response time. The stated 102-language and 73-language coverage should not be taken as proof of equal quality across languages.

The keyterm-library limit describes capacity, not its effectiveness for uncommon names, abbreviations or terms spoken with different accents. Nor does the announcement establish which language combinations benefit most from secondary-language selection or multilingual script output. Teams still need to validate their own audio, especially for overlapping speech, accents, noise and specialized terminology, and assess whether later turn commits are acceptable in their interaction design.

## Skill to build now

Create a small comparison set from your actual workload: recorded clips with mixed languages and live clips containing natural mid-sentence pauses. Run the recorded set with `scribe_v3`, selecting `numbers_format` and relevant `secondary_language_codes`; run the live set with `scribe_v3_live` and compare default commits against `commit_strategy=turn_prediction`. Include names, product terms and numbers in the test set, then try the keyterm library for the terms whose errors would break a downstream workflow. Review both accuracy and when a committed segment arrives.

## Reader takeaway

Scribe v3 targets recorded audio, while Scribe v3 Live adds controls designed for live, multilingual transcription. Treat the FLEURS improvements as a reason to run a comparison, not as a substitute for testing your own recordings, vocabulary and latency requirements.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Models](https://newruntime.com/topics/models.md): Explore the models topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)
