ElevenLabs launches Scribe v3 for recorded and live transcription

ElevenLabs launched Scribe v3 and Scribe v3 Live on November 10, 2026, giving developers separate models for recorded audio and live transcription. Both are available through ElevenAPI and are now defaults in the ElevenLabs app. Live is also available in ElevenAgents, according to the launch announcement.

The practical question is how the transcript enters your product. A meeting archive can wait for a completed recording. A voice agent needs to decide when a person has finished speaking, while a live captioning tool may need to revise which speaker said an earlier word. Those are different integration problems.

Choose the model for the audio workflow

Use the model ID scribe_v3 for recorded audio and scribe_v3_live for streaming. The transcription overview lists word timestamps, speaker labels, terminology prompting and output controls. Scribe transcribes speech in its original language rather than translating it.

The batch model supports prompts containing up to 1,000 keyterms and controls for writing numbers as digits or words. Live adds multilingual handling for speech that changes language within a sentence. The documentation says keyterm prompting and transcript editing incur extra charges, so a base transcription rate should not be treated as the complete cost of an enriched workflow. We have not verified a current plan-specific price.

Teams considering local processing should compare deployment requirements separately. Our Nemotron 3.5 Transcribe guide covers a recorded-audio model with a different deployment approach. That is a useful alternative to investigate for archives, but it is not a substitute for a streaming service.

Live speaker labels need an update path

The realtime diarization guide makes an important detail explicit: speaker labeling is off by default. Enabling diarize=true produces provisional labels in partial transcripts and a separate diarized event after each committed transcript. Labels can remain unresolved or change later.

Match the diarized result to the committed segment using segment_index. The service can also issue a speakers_merged event when two IDs turn out to represent the same speaker. Applications need to update earlier labels and handle successive merges in order. The guide says diarization cannot be combined with background-audio filtering.

For a meeting product, the consequence is architectural: do not store the first speaker label as an immutable fact. Keep a revision path for the displayed transcript and for any summary derived from it. Otherwise a later correction may fix the screen while leaving an action item attributed to the wrong person. Speaker separation also should not be presented as proof of a person’s identity.

Turn prediction is an explicit setting

The commit-strategy documentation says the default is manual, including for Scribe v3 Live. Set commit_strategy=turn_prediction to enable contextual turn detection. The model then estimates whether the speaker has finished, instead of relying only on a fixed silence period.

Silence detection remains active in this mode. The documented server behavior commits when the predicted completion probability reaches 0.9 or when the configured silence threshold is reached. That means the silence setting still matters when testing pauses, hesitations and long numbers. The guide also notes that SDK enums do not yet include the new strategy and describes the direct WebSocket setting. Check the SDK version and supported options before copying an example into production.

A sensible acceptance test should include people stopping halfway through an address, correcting themselves and resuming after a pause. Measure premature replies and time to a useful final transcript separately. An apparently faster response can be a worse experience if it answers before the caller has finished.

Read the accuracy claims in context

In its November 10 announcement, ElevenLabs reports a 27% relative reduction in word error rate for Scribe v3 against v2. For Live against v2 Realtime, it reports 43% lower semantic word error rate and 31% lower ordinary word error rate. These are vendor evaluations, not ByteForward tests.

The multilingual evaluation uses an internal FLEURS version across 102 languages, averaging languages equally. The launch also says its realtime speaker-scoring method differs from the batch method, so those diarization results should not be compared directly. The current overview lists 44% for Live’s semantic improvement; the launch says 43%. We retain the dated launch figure rather than treating the two as identical.

A relative reduction is not a percentage-point improvement. More importantly, an average across languages does not predict accuracy for your microphone, accent mix, jargon or overlapping speech. Build a small consented test set from the conditions the product actually serves, then score names, numbers and speaker attribution as well as overall text accuracy.

Check retention before sending sensitive audio

ElevenLabs’ Zero Retention Mode documentation describes an enterprise option for eligible API requests. It is not the general default, and web-interface traffic is explicitly outside its coverage. Account and support communications are also separate from the eligible request payloads.

Treat privacy configuration as its own rollout check. Confirm the account’s eligibility, the endpoint and the settings in use before sending confidential recordings. A successful transcription request proves that audio processing worked; it does not prove that the intended retention policy applied.

For an existing Scribe integration, changing the model ID is the starting point. A worthwhile migration also checks speaker revisions, turn settings, paid extras and retention, then compares the resulting workflow on representative audio. Those checks will tell a team more about the upgrade than a single launch percentage.