Notes on film, new tools and the decisions behind the work.

Tools / Analysis

ElevenLabs’ Scribe v3 listens for the end of a turn

The new transcription models target recorded audio and live conversation, with different accuracy claims and distinct tests behind them.

A sculpted metallic waveform with coloured lines. Conceptual Google Flow illustration; the waveform is decorative, not measured transcription data.
DU / SIGNALS · S/115Conceptual illustration · Google Flow; waveform is not measured data

A transcript can get every word right and still make a conversation hard to follow: speakers overlap, a name is unfamiliar, or a long pause is mistaken for the end of a thought. ElevenLabs’ Scribe v3 release, announced on 10 November, addresses recorded audio and live transcription as two related but different problems.

The company reports improvements against its previous models. The headline percentages use different error measures, however, and should not be read as a single ranking.

What the percentages actually measure

For Scribe v3, ElevenLabs reports a 27% relative reduction in word error rate versus Scribe v2 on its internal FLEURS benchmark covering 102 languages. That is a reduction in errors, not a claim that the model transcribes 27% more words correctly. FLEURS contains short, read speech; it does not by itself establish performance in a noisy meeting.

The distinction between relative and percentage-point changes is easy to miss. As a hypothetical example, reducing an error rate from 10% to 7.3% is a 27% relative reduction; it is a drop of 2.7 percentage points, not 27. Those example rates illustrate the arithmetic and are not ElevenLabs’ measured scores.

For Scribe v3 Live, the launch post reports a 43% reduction in semantic word error rate versus v2 Realtime, alongside a 31% reduction in ordinary word error rate. Semantic WER assesses meaning differently from ordinary WER, so 43 and 27 are not directly comparable. The documentation headline says 44% for the semantic figure, while the launch post says 43%; ElevenLabs has not explained the discrepancy in those sources.

ElevenLabs says it scored each language separately and then averaged the results, giving every language equal weight. A single pooled score from another evaluation can therefore answer a different question, even when both are called WER. The company also built a separate dataset of meetings, agent calls and voice notes for speaker evaluation because FLEURS clips contain one speaker.

Sources: ElevenLabs [1] · ElevenLabs [2]

X / DEMO

ElevenLabs introduces Scribe v3

Loading post from X…

Credit / ElevenLabsOpen original 

Official ElevenLabs launch post.

Words and speakers are separate problems

Both models also identify speakers. For the recorded-audio Scribe v3, ElevenLabs reports fewer diarization errors—mistakes in assigning speech to a person—on its dataset of meetings, agent calls and voice notes, including a 17% reduction for clips with three or more speakers compared with v2. This is a company-reported result on its own dataset, not an independent guarantee for every recording.

Speaker labels answer ‘who spoke when’; word error rate asks whether the words match a reference. Improving one does not prove the other. For an interview, both matter: an accurate sentence attributed to the wrong person can still mislead a reader.

The live speaker test needs another qualification. ElevenLabs aligned each transcript to a reference before scoring which speaker received each word, specifically to isolate speaker assignment from missed words. The company says that procedure lowers all models’ error rates; its live figures should therefore not be compared directly with the batch diarization figures.

Sources: ElevenLabs [1]

A pause is not always the end

Scribe v3 Live adds turn prediction: the system considers conversational context when deciding whether a speaker has finished. That distinction matters to a voice assistant waiting to respond. Interrupt too early and it cuts across a thought; wait too long and the exchange feels slow.

It is a model prediction, not a promise that pauses and interruptions will always be interpreted correctly. The useful test for a team is its own speech: long pauses, interruptions, accents and names from the actual work. We have not run that test for this article.

Sources: ElevenLabs [1] · ElevenLabs [2]

Teach the model your vocabulary

Keyterm prompting lets users supply vocabulary such as names and specialist terms. ElevenLabs’ documentation lists up to 1,000 keyterms for batch Scribe v3; the Live documentation describes a larger library but gives inconsistent limits, so its exact capacity needs confirmation. Keyterm prompting and transcript edits can carry additional charges.

For a recorded interview or archive, word-level timestamps offer a route back from a line of text to the audio that needs checking. ElevenLabs also documents an experimental transcript-editing option that returns edited text alongside the original, keeping the original timestamps and speaker labels. A cleaner version can help with formatting; it should not be mistaken for a verbatim record of every hesitation.

ElevenLabs lists both models in its API and says the app uses the new release. A multilingual transcript still requires a human review where wording and attribution matter; transcription itself is not translation.

Sources: ElevenLabs [1] · ElevenLabs [2]

SOURCES & METHOD

The record behind the article

Checked 11 November 2026 against ElevenLabs’ launch announcement and documentation. The reported figures are the company’s; we have not independently tested the models. Cover: A sculpted metallic waveform with coloured lines. Conceptual Google Flow illustration; the waveform is decorative, not measured transcription data.

  1. Introducing Scribe v3 ElevenLabs · 2026-11-10
  2. Speech to text ElevenLabs
  3. Scribe v3 launch post ElevenLabs