ElevenLabs Scribe v3: Power-User Guide to Scribe v3 Live, Pricing and the API

Last updated: November 14, 2026

ElevenLabs Scribe v3 is the company’s newest speech to text model, announced November 10, 2026, with a real-time sibling called Scribe v3 Live. ElevenLabs reports 27% fewer word errors than Scribe v2, better speaker labels and new formatting controls. API prices are $0.20 per hour for Scribe v3 and $0.22 per hour for Live.

This guide covers what changed, how to upgrade, the new parameters, pricing and the limits to know. Details come from the ElevenLabs launch announcement, the models documentation, the API changelog and the API pricing page.

Contents

What is new in Scribe v3

ElevenLabs released two models. scribe_v3 transcribes recorded audio through the batch endpoint, POST /v1/speech-to-text. scribe_v3_live transcribes streaming audio through the realtime WebSocket at wss://api.elevenlabs.io/v1/speech-to-text/realtime. Both are available in ElevenAPI, Scribe v3 Live is also available in ElevenAgents, and the announcement says both are now the default in the ElevenLabs app.

According to the announcement, the release responds to five recurring customer requests: better diarization with three or more speakers, turn detection that does not cut off callers who pause, realtime accuracy closer to batch, stronger Hindi and Hindi-English code switching, and more control over output formatting.

The models page states that on the FLEURS benchmark across 102 languages, Scribe v3 lowers word error rate (WER) from 16.94% to 12.38%. ElevenLabs computes FLEURS WER per clip and averages per language, so it warns that these figures are not comparable with corpus-level results published elsewhere.

For Scribe v3 Live, the announcement reports 43% lower semantic WER than Scribe v2 Realtime across 102 languages, while the models page and changelog give 44% across 73 languages. The models page also lists WER on Earnings-22, accented English financial calls, falling from 5.99% to 3.91%. All accuracy figures are ElevenLabs’ own.

Scribe v3 compared with Scribe v2

The table summarizes the main differences listed in the official documentation and pricing page.

Item Scribe v3 Scribe v3 Live Scribe v2 / v2 Realtime
Model ID scribe_v3 scribe_v3_live scribe_v2 / scribe_v2_realtime
API price $0.20 per hour $0.22 per hour $0.20 / $0.22 per hour
Speaker diarization Up to 32 speakers; 17% fewer errors than v2 with 3+ speakers New: realtime diarization with diarize=true Batch only
Keyterms Keyterm prompting, up to 1,000 terms 50 standard keyterms plus a keyterm library of up to 10,000 terms Keyterm prompting
New controls numbers_format, secondary_language_codes Turn prediction, multilingual mode None of these
Languages 90+ 90+, automatic detection 90+

Scribe v3 keeps the Scribe v2 feature set, including word-level timestamps, audio event tagging, entity detection, transcript editing, a no verbatim mode, export to SRT, DOCX and PDF, and webhooks for asynchronous results.

How to upgrade and use the new features

ElevenLabs describes the upgrade as a one-line change. Existing request parameters keep working, and Scribe v2 stays available if you need to roll back.

  1. Change the batch model. Replace scribe_v2 with scribe_v3 as model_id on POST /v1/speech-to-text.
  2. Set number formatting if needed. Add numbers_format with digits (42) or words (forty-two). If you omit it, the model chooses.
  3. Declare mixed languages. Set language_code to the main language and repeat the secondary_language_codes form field once per extra language, using ISO-639-1 or ISO-639-3 codes. A JSON array sent as one string is not accepted.
  4. Pass new fields carefully in SDKs. The number and language options guide says the SDKs do not yet have typed arguments for these two parameters. In Python, pass them through request_options; in TypeScript, send the request with fetch.
  5. Switch realtime traffic. Replace scribe_v2_realtime with scribe_v3_live as the model_id query parameter. The default commit strategy remains manual, so behavior does not change until you opt in to turn prediction.

Both new batch parameters work only with scribe_v3; other models return an invalid_parameters error. Unrecognized language codes are skipped without an error, so check them against the supported language list. Scribe v3 transcribes each language as spoken and does not translate.

Turn prediction in Scribe v3 Live

Voice agents usually decide that a caller has finished speaking after a fixed silence. A short threshold interrupts people who pause mid-sentence; a long one makes the agent feel slow. ElevenLabs says Scribe v3 Live uses conversational context as well as audio, so a caller pausing halfway through a number is less likely to be cut off.

According to the commit strategies guide, every partial_transcript message from Scribe v3 Live includes turn_probability. With commit_strategy=turn_prediction, the server commits the segment when that value reaches 0.9 or higher. VAD stays active, so a segment is also committed when the silence threshold is reached. Other models reject this strategy with an invalid_request error.

The guide notes that the Python and JavaScript SDK enums do not include turn_prediction yet, and the models page says the same for multilingual mode and the keyterm library. To use them, connect to the WebSocket directly with a URL such as ?model_id=scribe_v3_live&audio_format=pcm_16000&commit_strategy=turn_prediction. With manual commits, you can still read turn_probability and decide in your own code.

ElevenLabs cites Eneco, a Dutch energy supplier, whose voice agent recognized postcodes correctly on the first try 40% more often in testing. If you are building voice agents on other platforms, compare our guides to Cursor Voice Chat and the DeepSeek application APIs.

Limits and caveats

  • File limits. The speech to text documentation lists files up to 5 GB and up to 15 hours of audio in standard mode.
  • Vendor benchmarks. Accuracy figures come from ElevenLabs’ own tests, and the announcement and documentation report slightly different realtime numbers. Test on your own audio before switching production traffic.
  • SDK gaps. Several new features need raw HTTP form fields or a direct WebSocket connection until the SDKs catch up.
  • Keyterm limits differ. Batch keyterm prompting takes up to 1,000 terms; realtime takes 50 standard keyterms plus the larger library.
  • Enterprise options. The announcement says both models run in the US by default, with isolated EU, India and Singapore environments for enterprise data residency, Zero Retention Mode, and HIPAA eligibility with a Business Associate Agreement.

For other recent speech and voice models, see our Nemotron 3.5 guide, which covers NVIDIA’s Transcribe model, the ZONOS3-Mini voice cloning guide and Gemini Voice Remixing.

Frequently asked questions

What is ElevenLabs Scribe v3?

Scribe v3 is the newest ElevenLabs speech to text model for recorded audio, announced November 10, 2026. Its live counterpart, Scribe v3 Live, handles real-time audio. Both replace Scribe v2 and Scribe v2 Realtime as the most accurate options, and the older models remain available.

How much does Scribe v3 cost?

The ElevenLabs API pricing page lists Scribe v3 at $0.20 per hour of audio and Scribe v3 Live at $0.22 per hour. Those are the same rates listed for Scribe v2 and Scribe v2 Realtime, so upgrading does not change the listed price.

How do I switch from Scribe v2 to Scribe v3?

Change the model ID. For recorded audio, replace scribe_v2 with scribe_v3 on POST /v1/speech-to-text. For live audio, replace scribe_v2_realtime with scribe_v3_live on the realtime WebSocket. ElevenLabs says existing request parameters work without changes.

What is turn prediction in Scribe v3 Live?

Turn prediction commits a transcript segment when the model estimates the speaker has finished their turn, rather than at the first pause. Set commit_strategy=turn_prediction; the server commits when turn_probability reaches 0.9 or higher. The default strategy is still manual.

Does Scribe v3 support speaker diarization?

Yes. Scribe v3 supports diarization for up to 32 speakers, and ElevenLabs reports 17% fewer diarization errors than Scribe v2 on recordings with three or more speakers. Scribe v3 Live adds realtime diarization for the first time with diarize=true.

Sources

Related answers

Leave a Comment