ElevenLabs Scribe v3: What Interview Transcription Changes Mean for Research Ethics

ElevenLabs announced Scribe v3 and Scribe v3 Live, new speech-to-text models, on 10 November 2026. For researchers who transcribe interviews and focus groups, the changes that matter are speaker labelling (diarization), language coverage and the way ElevenLabs documents data handling. Everything below is attributed to ElevenLabs; the accuracy figures are the vendor’s own benchmark results, and this post has not found independent verification of them.

What ElevenLabs released

• Scribe v3 (API model ID scribe_v3) is for recorded audio, the batch case most interview studies use.
• Scribe v3 Live (scribe_v3_live) is for low-latency real-time transcription.
• ElevenLabs’ blog post, dated 10 November 2026, also describes a version of Scribe v3 Live optimized for medical conversations, aimed at healthcare teams using its ElevenAgents product.
• ElevenLabs’ developer changelog lists the release under 9 November 2026 and states that Scribe v2 and Scribe v2 Realtime remain available. The one-day difference between the changelog and the blog is a publication-date discrepancy between two ElevenLabs pages, not a separate event.

Languages, speaker labels and accuracy, as ElevenLabs states them

• Languages. ElevenLabs says Scribe v3 "makes 27% fewer word errors than Scribe v2 across 102 languages" on the FLEURS benchmark, with per-language scores averaged so each language counts equally. It reports larger reductions for some languages, for example German (42%) and Spanish (35%), and 23% for English.
• Diarization. ElevenLabs says Scribe v3 has 12% fewer diarization errors than Scribe v2 overall and 17% fewer on recordings with three or more speakers, which is the focus-group case. It says Scribe v3 Live adds diarization to real-time transcription for the first time.
• Live model. ElevenLabs reports 43% fewer errors on semantic word error rate versus Scribe v2 Realtime (its changelog gives 44% across 73 languages, so the two ElevenLabs pages are not identical on this figure).
• Other documented features. The changelog lists a numbers_format option, secondary_language_codes for audio that mixes languages, and a keyterm library of up to 10,000 terms for the live model.
• Timestamps. The launch post does not discuss timestamps, so this post makes no claim about them; check the current API documentation if time-aligned output matters to your coding workflow.

ElevenLabs cautions that its results may not be directly comparable with figures measured differently. Benchmark gains on read speech or curated datasets do not guarantee the same performance on your recordings, which may include dialect, background noise, overlapping talk or specialist vocabulary.

What research ethics and IRB reviewers will ask

Sending interview audio to a cloud transcription service is a data-processing decision that research ethics boards typically expect to see described in the protocol and consent materials. ElevenLabs documents the following:

• Processing location. Per the launch post, Scribe v3 and v3 Live run in the US by default, and enterprise customers with data residency requirements can use isolated environments in the EU, India and Singapore. This is an enterprise option, so a standard account should be assumed to process in the US unless ElevenLabs confirms otherwise in writing.
• Retention. The launch post describes Zero Retention Mode, which deletes audio and text as soon as a request completes. ElevenLabs’ documentation says it applies to eligible API endpoints, including those beginning /v1/speech-to-text/, not to the web interface or to data sent outside API payloads such as support tickets, and that for API use it is enabled per request with enable_logging=false. Using the web app is therefore a different data-handling situation from the API.
• Security and compliance programs. ElevenLabs states both models are covered by its SOC 2, ISO 27001 and PCI DSS Level 1 programs, and are HIPAA-eligible with a Business Associate Agreement. It also offers private deployment.

What the sources reviewed here do not establish: whether a given institution’s data-protection office will accept these terms, how consent wording should change, or any ElevenLabs commitment specific to research data beyond the points above. Those require the contract, the data processing agreement and local review.

What it means for research offices

• Treat the tool as a third-party processor in the data management plan: record the vendor, processing region, retention setting and who holds the account.
• Where participants were promised that recordings stay on institutional systems, a cloud model needs a protocol amendment and updated consent, not just a tool swap.
• Diarization labels are machine output. Researchers should verify attribution of speech to participants against the audio before analysis, particularly where identity or sensitive content is involved.
• Keep the model ID and version in the methods record, since moving from Scribe v2 to v3 can change transcripts for the same audio.

Sources

• ElevenLabs, Scribe v3, blog post, 10 November 2026.
• ElevenLabs, Developer changelog, Scribe v3 entry, 9 November 2026.
• ElevenLabs, Zero Retention Mode documentation, accessed 13 November 2026.
• Runtime Wire, ElevenLabs launches Scribe v3 and Scribe v3 Live for messier real-world audio, 10 November 2026 (headline and byline only; article body not reviewed).