Overview
Scribe v3 is the pair of speech-to-text models ElevenLabs released on 10 November 2026: Scribe v3 (`scribe_v3`) transcribes recorded audio and Scribe v3 Live (`scribe_v3_live`) transcribes realtime audio. They succeed Scribe v2 and Scribe v2 Realtime, and ElevenLabs says upgrading takes only a model ID change. ElevenLabs' docs list both at 90+ languages, and both became the default in the ElevenLabs app.
For recorded audio, ElevenLabs reports 27% fewer word errors than Scribe v2 on FLEURS across 102 languages: average WER falls from 16.94% to 12.38%, and the docs say it improves on every language tested. Diarization handles up to 32 speakers, and ElevenLabs says it makes 12% fewer diarization errors than v2 overall and 17% fewer on recordings with three or more speakers. Other features are keyterm prompting with up to 1,000 terms, numbers as digits or words, secondary languages for mixed-language audio, entity detection, word-level timestamps, and exports to formats such as SRT, DOCX and PDF.
Scribe v3 Live is built for voice agents and live captioning. ElevenLabs reports 43% fewer errors than Scribe v2 Realtime, measured as semantic WER (an error counts only when the meaning changes), and 31% lower standard WER. It adds speaker diarization to the realtime model for the first time, plus turn prediction from conversation context so a caller who pauses mid-sentence is not cut off. It also adds a multilingual mode for mid-sentence language switches and a keyterm library of up to 10,000 terms. ElevenLabs' product page quotes under 150 ms latency, and a version tuned for medical conversations ships in ElevenAgents.
All of the comparisons are ElevenLabs' internal tests. FLEURS is a public benchmark, but the diarization results come from the company's own dataset of real meetings, agent calls and voice notes, which it says it will open-source. On its API pricing page, ElevenLabs lists batch transcription at $0.20 per hour of audio and realtime at $0.22 per hour.
| Released | 2026-11-10 |
|---|---|
| License | Proprietary |
| Weights | API only |
| Parameters | Not disclosed |
| Modalities | Audio, Text |
| Status | Generally available |
Benchmarks
ElevenLabs' batch (non-streaming) comparison from the Scribe v3 launch post, November 2026. FLEURS rows are averages over all 102 FLEURS languages and were run only for the models shown with a score. Diarization error rate (DER) comes from ElevenLabs' own dataset of real recordings in 16 languages. Lower is better on every row.
| Benchmark | Scribe v3 | Scribe v2 | MAI-Transcribe-2 | Muse Voice Transcribe | AssemblyAI Universal-3.5 Pro | Gemini 3.5 Transcribe | Deepgram Nova-3 |
|---|---|---|---|---|---|---|---|
| FLEURS WER · 102 languages | 12.4% | 17% | — | — | — | 18.5% | — |
| FLEURS semantic WER · 102 languages | 4.4% | 5.4% | — | — | — | 7.5% | — |
| Diarization error rate · all recordings | 5.6% | 6.4% | 8.5% | 8.8% | 10.4% | 10.6% | 14.2% |
| Diarization error rate · 3+ speakers | 8% | 9.6% | 15% | 10.6% | 13.7% | 14.3% | 18.4% |
Pricing
| Input | $0.20–$0.22 / hour of audio |
|---|
Scribe v3 (batch) $0.20 per hour; Scribe v3 Live (realtime) $0.22 per hour on the ElevenLabs API, with included hours on each plan.
Strengths
- 27% fewer word errors than Scribe v2 on FLEURS across 102 languages (12.4% vs 17.0% WER), per ElevenLabs
- Lowest diarization error among the non-streaming models in ElevenLabs' test: 5.6% DER, against 6.4% for Scribe v2 and 8.5% to 14.2% for five other providers' models
- Scribe v3 Live reports 3.1% English diarization error against 5.9% to 7.3% for the other streaming models ElevenLabs tested
- Turn prediction from conversation context and a multilingual mode for mid-sentence language switches in the realtime model
- Keyterm prompting with up to 1,000 terms (batch) and a keyterm library of up to 10,000 terms (Live)
- Drop-in upgrade from Scribe v2: change the model ID and existing request parameters work unchanged
Best for
- Reach for Scribe v3 for meeting, interview and podcast transcripts that need accurate speaker labels for up to 32 speakers.
- Reach for Scribe v3 Live as the listening stage of a voice agent, where turn prediction keeps a pausing caller from being cut off.
- Reach for keyterm prompting when audio is full of product names, postcodes or jargon that a general model would mishear.
- Reach for the multilingual mode for callers who switch language mid-sentence.
How to access
| Provider | Model ID |
|---|---|
| ElevenLabs API — Scribe v3 ↗ | scribe_v3 |
| ElevenLabs API — Scribe v3 Live ↗ | scribe_v3_live |
FAQ
What is Scribe v3?
Scribe v3 is ElevenLabs' speech-to-text model for recorded audio, released on 10 November 2026 alongside Scribe v3 Live for realtime audio. Both are on the ElevenLabs API as `scribe_v3` and `scribe_v3_live`, and they succeed Scribe v2 and Scribe v2 Realtime.
How much more accurate is Scribe v3 than Scribe v2?
On FLEURS across 102 languages, ElevenLabs reports 27% fewer word errors, with average WER falling from 16.94% to 12.38%. For Scribe v3 Live it reports 43% fewer errors than Scribe v2 Realtime on semantic WER and 31% lower standard WER. These are ElevenLabs' own tests.
What does Scribe v3 cost?
ElevenLabs' API pricing page lists Scribe v3 at $0.20 per hour of audio and Scribe v3 Live at $0.22 per hour, with a number of included hours on each subscription plan.
Does Scribe v3 Live label speakers?
Yes. Speaker diarization is new to the realtime model in v3. ElevenLabs reports a 3.1% diarization error rate on English recordings, against 5.9% to 7.3% for the other streaming models it tested. Draft speaker labels appear in partial transcripts.
How do I upgrade from Scribe v2?
Change `model_id` from `scribe_v2` to `scribe_v3`; ElevenLabs says existing request parameters work unchanged. For Scribe v3 Live the default commit strategy is still `manual`, so turn prediction is used only once you set `commit_strategy` to `turn_prediction`.
How many languages does Scribe v3 support?
ElevenLabs' docs list 90+ languages for both Scribe v3 and Scribe v3 Live. The headline FLEURS accuracy figures for Scribe v3 are averaged over all 102 FLEURS languages.