Introduction

A voice agent is only as good as what it hears. If it mishears one digit of a postcode, or cuts you off while you stop to read out a card number, the call falls apart, no matter how lovely the voice sounds. ElevenLabs went after exactly that problem today.

On November 10 (UTC), the day before its NYC Summit, ElevenLabs launched Scribe v3 and Scribe v3 Live, its new speech-to-text models. Scribe v3 handles recorded audio. Scribe v3 Live is the low-latency version for streaming. Both are closed, API-only models, available now in ElevenAPI and as the default in the ElevenLabs app. Scribe v3 Live also powers ElevenAgents. Upgrading is a one-line change to the model ID, and the price per hour hasn't changed.

What Shipped

  • Models: scribe_v3 (batch, on POST /v1/speech-to-text) and scribe_v3_live (realtime, on the wss://api.elevenlabs.io/v1/speech-to-text/realtime WebSocket).
  • Vendor: ElevenLabs, announced in a blog post and on
    .
  • Modality: speech-to-text in 90+ languages, with speaker diarization, word-level timestamps and audio event tags.
  • Access: closed weights. Use it through the API, the ElevenLabs app or ElevenAgents. Enterprises can also get private deployment, plus isolated EU, India and Singapore regions.
  • Price: $0.20 per hour for Scribe v3 and $0.22 per hour for Scribe v3 Live, the same as Scribe v2 and v2 Realtime. Keyterm prompting costs $0.05 per hour extra, and transcript editing costs $0.07 per hour.
  • Compliance: Zero Retention Mode, SOC 2, ISO 27001, PCI DSS Level 1, and HIPAA eligibility with a BAA.

The Numbers

ElevenLabs scores each language separately and then averages them, so every language counts equally. The company notes this gives higher WER numbers than the pooled scores most vendors publish, so don't compare these figures directly with other leaderboards.

  • Scribe v3 vs Scribe v2 (FLEURS, 102 languages): average WER drops from 16.94% to 12.38%, which is 27% fewer word errors. It improved in every language tested.
  • Biggest batch gains: German 42%, Spanish 35%, English 23%, Swahili 34%, Tamil 29% and Thai 28%.
  • Diarization: 12% fewer speaker-label errors than Scribe v2 overall, and 17% fewer on recordings with three or more speakers.
  • Scribe v3 Live vs Scribe v2 Realtime: the launch post says 44% lower semantic WER, while the blog says 43% semantic WER and 31% plain WER across 102 languages. On Earnings-22 (accented English financial calls), WER falls from 5.99% to 3.91%.
  • Biggest realtime gains: Greek 77%, Danish 54%, Croatian 45% and Hindi 36%.
  • Realtime diarization: a 3.1% diarization error rate on the English part of ElevenLabs' own real-world set, compared with 5.9% to 7.3% for the other realtime models it tested. ElevenLabs says it will open-source this 16-language set of meetings, agent calls and voice notes.

Semantic WER counts an error only when the meaning changes, so "42" and "forty-two" aren't marked wrong. That's a fairer test for streaming, where models format text before the sentence ends. Still, every number here comes from ElevenLabs itself, so test it on your own audio.

What's New for Builders

Scribe v3 (recorded audio)

  • numbers_format: choose digits (42) or words (forty-two).
  • secondary_language_codes: list the other languages you expect in mixed-language audio.
  • Everything from v2 stays: up to 32 speakers, keyterm prompting with up to 1,000 terms, entity detection, plain-English transcript edits like "write every date as YYYY-MM-DD", SRT/DOCX/PDF export and webhooks.

Scribe v3 Live (streaming)

  • Turn prediction: with commit_strategy=turn_prediction, the server commits a segment when the model thinks the speaker has finished, using the conversation and not just a silence timer. If a caller pauses halfway through a phone number, they don't get cut off.
  • Live diarization: with diarize=true, partial transcripts carry draft speaker labels, and a separate diarized_transcript event follows each committed segment. Labels can change later through speakers_merged events.
  • Multilingual mode: multilang_mode=true writes each language in its own script, even when the speaker switches mid-sentence. Hindi–English callers were a big reason for this.
  • Keyterm library: a large vocabulary the model checks only when it hears a likely match. The docs list up to 10,000 terms, on top of the 50 standard keyterms.

In a pre-launch test, Dutch energy company Eneco ran Scribe v3 Live on its customer calls. Its agent got postcodes right on the first try 40% more often, and house numbers 15% more often. For healthcare teams, ElevenLabs is also releasing a medical version of Scribe v3 Live inside ElevenAgents.

How to Try It

For recorded audio, just swap the model ID:

curl -X POST https://api.elevenlabs.io/v1/speech-to-text \
  -H "xi-api-key: $ELEVENLABS_API_KEY" \
  -F model_id=scribe_v3 \
  -F diarize=true \
  -F numbers_format=digits \
  -F file=@meeting.mp3

For streaming, connect to the WebSocket directly and add the new options as query parameters:

wss://api.elevenlabs.io/v1/speech-to-text/realtime?model_id=scribe_v3_live&audio_format=pcm_16000&commit_strategy=turn_prediction&diarize=true

Send 16 kHz mono 16-bit PCM in chunks of 0.1 to 1 second.

Pro Tip

Read the fine print before you build. The models page says the official SDKs don't support turn prediction, multilingual mode or the keyterm library yet, so use the raw WebSocket for those, even though the blog shows an SDK snippet. Live diarization also can't be combined with filter_background_audio. The default commit strategy is still manual, so existing integrations behave exactly the same until you opt in.

Where It Fits

Scribe v3 is for meeting notes, captions with speaker labels, podcast and interview transcripts, and contact-center QA. Scribe v3 Live is for voice agents, live captions and dictation, especially where callers read out numbers or switch languages. If you need open weights you can run yourself, NVIDIA's Nemotron 3.5 Transcribe is the one to compare. Fish Audio's transcribe-stream-1 preview aims at the same turn-taking problem. ElevenLabs' own Eleven Realtime speech-to-speech models are a separate product line, and Scribe v3 Live doesn't replace them.

Conclusion

Scribe v3 is a quiet upgrade, but a useful one. You get fewer wrong words, speaker labels you can trust in a crowded meeting, and, for the first time, live diarization plus turn-taking that waits while you finish reading out a number. It costs the same per hour as v2, and switching takes one line. Run your hardest audio through it, the overlapping meetings and mid-sentence code-switching, and see how the numbers hold up. I'll be watching what else ElevenLabs brings to the Summit stage.

—Anabel ♡