Title card reading 'Scribe v3 Live Medical' with the subtitle 'a clinical transcriber you cannot call'. Below it a simple flow diagram: a phone handset labelled 'patient call' feeding a rounded box labelled 'ElevenAgents', with a greyed-out and struck-through box beside it labelled 'Speech to Text API'.
Guides & Insights

Scribe v3 Live Medical: ElevenLabs Built a Clinical Transcriber You Cannot Call

Author

jinhao.song

Date Published

Latest models · 20View all models →
Benchmarks: Artificial Analysis · updated daily
Back to all posts

The most interesting fact about Scribe v3 Live Medical is a deployment fact, not an accuracy number: there is no model ID. ElevenLabs shipped the model, documented it, benchmarked it on clinical audio, put it in front of healthcare customers — and then closed the only door most developers use. It runs inside ElevenAgents, selected from an ASR dropdown in an agent's settings, and it is explicitly not available through the Speech to Text API. The API model for medical audio remains Scribe v2 Medical, a batch model from September. So the thing ElevenLabs announced on 9 November 2026 as part of the Scribe v3 family launch — Scribe v3 and Scribe v3 Live for recorded and real-time audio, with the medical real-time fine-tune following into early access — is best understood as a platform feature wearing a model's name.

That framing decides whether the release matters to you. If you are building a clinical voice agent on ElevenLabs' agent runtime, the best real-time medical ASR the company has is now one dropdown away. If you are building that agent anywhere else, this release changed nothing about your options, and the benchmark chart ElevenLabs published is a chart about a model you cannot rent.

What actually shipped, and when

ElevenLabs' changelog dates the family launch to 9 November 2026: scribe_v3 for recorded audio, replacing scribe_v2 on the existing transcription endpoint, and scribe_v3_live on the realtime WebSocket. Both are documented as 27% and 44% relative improvements respectively — Scribe v3 lowers word error rate from 16.94% to 12.38% on FLEURS, and Scribe v3 Live claims a 44% lower semantic WER than Scribe v2 Realtime. Neither figure is a medical figure. They are the general-purpose numbers for the family the medical variant is built on.

The medical variant appears separately, in the model documentation and on ElevenLabs' medical transcription page, with three attributes that would be unusual on a launch post and are ordinary on a platform feature:

• Availability — ElevenAgents only. Not on the Speech to Text API, no model ID to pass.

• Status — early access, gated to workspaces that have it enabled, and it requires the company's turn_v3 turn model.

• Cost — no separate rate. ElevenLabs says it is included in standard ElevenAgents pricing, with up to 10,000 custom terms at no extra charge.

Compare that with Scribe v2 Medical, which is a normal API product: pass scribe_v2_medical as a model ID, pay $0.20 per hour, same rate as Scribe v2. The v3 medical model did not get that treatment. It got a dropdown.

A section of the ElevenLabs medical transcription page headed 'Medical transcription built for live calls and recorded audio', with the line 'Two Speech to Text models for clinical language: Scribe v3 Live Medical for ElevenAgents, and Scribe v2 Medical for batch transcription', and a sample transcript reading 'He's on metformin 1000 twice a day, empagliflozin 10, and atorvastatin 40. A1C 8.4, eGFR 52, LDL 128'.

The claim that matters is a clinical one, on a small public set

ElevenLabs' headline for Scribe v3 Live Medical is not a FLEURS number. It is PriMock57 — 57 mock primary-care consultations recorded for research, which is small, public, and much closer to the acoustics of an actual clinical call than read speech is. On that set ElevenLabs reports the lowest word error rate among real-time models it tested: 12.8% for Scribe v3 Live Medical, against 15.5% for AssemblyAI's Universal-3.6 Pro Realtime in medical mode, 16.3% for Speechmatics Oak-1, 17.0% for Deepgram Nova-3 Pharma and 17.2% for Nova-3 Medical.

Those are vendor-run numbers on a vendor-selected benchmark, and the honest reading is narrow: a 2.7-point gap to the nearest competitor on one 57-consultation set is a real signal that a clinical fine-tune helps, and it is not a guarantee that your cardiology clinic's telephony audio will follow the same ranking. Nobody outside ElevenLabs has reproduced it, and the set is small enough that per-call variance is larger than the gap.

The more useful disclosure is buried in the medical transcription page's own FAQ, and it is unusually candid: isolated drug names with no context are the hardest case for every model, and ElevenLabs' advice is to use keyterm prompting to bias recognition toward expected terms. That is a company telling you where its fine-tune stops. A drug name inside a sentence — "he's on metformin 1000 twice a day" — is a solvable problem, and the sample transcript on that page is built from exactly that shape. A drug name read out alone, as a caller might do when asked to repeat a prescription, remains hard for everyone.

A single-column scoreboard titled 'Scribe v3 Live Medical — the scoreboard', with rows reading 'Availability: ElevenAgents only, no model ID', 'Status: early access, requires turn_v3', 'PriMock57 WER: 12.8%', 'Nearest rival on that set: 15.5%', 'Price: no separate rate, bundled' and 'PHI: docs say do not send yet', with a footer reading 'All figures vendor-reported by ElevenLabs; no independent reproduction.'

Turn detection is the feature that makes it a medical model

The part of this release that will change how a clinical agent feels is not accuracy. It is turn prediction. General-purpose streaming ASR commits a transcript segment when it detects silence, which is why voice agents interrupt people reading out long strings — a member ID, a date of birth, a phone number, a dosage spelled out slowly. Scribe v3 Live Medical's turn model uses conversation context to decide when a caller has actually finished, so the agent waits through the pauses inside a member ID instead of replying to the first three digits.

That behaviour is a property of the turn_v3 model coupled to the ASR, and it is available on Scribe v3 Live generally, not only the medical fine-tune. What the medical variant adds is the vocabulary bias underneath it: the recognition of the clinical terms that were being mangled in the first place, over 90+ languages, with a multilingual mode that keeps up when a caller switches language mid-call. ElevenLabs says medical accuracy improvements were measured in English, French and German, which is a narrower claim than the 90+ language support and worth stating separately.

The compliance line is the one to read twice

ElevenLabs' medical transcription page says both models are HIPAA-eligible for Enterprise customers with a Business Associate Agreement in place and Zero Retention Mode enabled, and that with ZRM on, audio input and transcript output are not retained after processing. Its documentation page for Scribe v3 Live Medical says something firmer and less convenient: do not send protected health information to Scribe v3 Live Medical until HIPAA eligibility is confirmed for this model, and EU data residency is not available during early access.

Both sentences are on ElevenLabs' own site, and they are not the same sentence. The reconciled reading is that the general healthcare compliance posture — BAA, ZRM, SOC 2, GDPR — describes the platform and the GA medical product, Scribe v2 Medical, while the early-access real-time model has not finished its own eligibility work. For a covered entity, the documentation warning is the one that governs: an early-access model without confirmed HIPAA eligibility is not a place for PHI, whatever the marketing page's comparative table implies. That is a temporary state, and it is also the single fact most likely to stop a compliance review this quarter.

What it costs, honestly

Because Scribe v3 Live Medical has no meter of its own, there is no per-minute price to quote. The nearest published numbers are the ones the family carries on the API: Scribe v3 at $0.20 per hour of audio, Scribe v3 Live at $0.22 per hour, keyterm prompting at $0.08 per hour on the realtime models and $0.05 per hour on the batch ones. If you want a mental model for what the medical model would cost if it ever reached an API, those are the rates it would plausibly inherit — the same way Scribe v2 Medical inherited Scribe v2's $0.20.

What you actually pay today is an ElevenAgents subscription, which bundles the ASR with the agent's LLM, telephony, turn-taking and tool calls. That is a real difference in shape, not just in number: an agent platform prices per minute of conversation, and conversation includes the LLM tokens and the audio output, so the transcriber's share is not separable. Teams that model costs per ASR minute will find this bundle opaque; teams that were going to assemble an agent anyway will find it cheaper than the sum of parts.

Where the reasoning half of the agent comes from

A clinical voice agent is three models in a trench coat: something that hears, something that decides, and something that speaks. Scribe v3 Live Medical covers the first, and only inside ElevenAgents. The middle is the part you can route independently — the model that reads the transcript, checks a coverage rule, drafts the note, or decides the call needs a human. OrcaRouter puts 215 models behind one API at 0% markup, passing each provider's list price through unchanged, so the model that reasons over a transcript can be swapped, A/B'd or failed over without touching the ASR leg.

To be exact about what we do and do not serve: OrcaRouter does not host Scribe v3 Live Medical, Scribe v3 Live, or any ElevenLabs transcription model, and none of the rates above are ours. What we route is the LLM side of the pipeline — the Claude, Gemini, GPT, Qwen and DeepSeek models a clinical agent calls after the words arrive. Pricing for those is the provider's own, unmarked.

A section of the OrcaRouter site headed 'Zero markup. Zero black boxes.', with the line 'You pay each provider their exact list price. We add $0 per token' and further text explaining that cached and repeated prompt tokens bill at the provider's own cache rate and that cached_tokens appears on every receipt.

Who should act on this, and who should wait

If you are on ElevenAgents and taking patient, member or provider calls — scheduling, intake, refill requests, benefits questions — the case for switching the ASR dropdown is straightforward. You get clinical vocabulary bias, a turn model that waits out member IDs, and 10,000 custom terms, at no additional line item. The PriMock57 chart is the vendor's, but the direction of the improvement is the same one every clinical ASR vendor reports, and the feature list is specific enough to test.

If you are not on ElevenAgents, wait. Nothing here is callable from your stack, and the substitute ElevenLabs offers through the API is a batch model from September: Scribe v2 Medical, 15% fewer errors than Scribe v2 on the Eka Medical ASR test split, $0.20 per hour, keyterms and diarization included. For ambient documentation that is processed after the visit, that is a reasonable product and it is the one with a confirmed compliance story. For a live call, your options remain the general-purpose real-time models, with a phrase list, and the accuracy gap the medical fine-tune is claiming.

The thing to watch is not a benchmark update. It is the model ID. When scribe_v3_live_medical appears in the Speech to Text reference — or when the early-access warning about PHI disappears from the docs page — this becomes a launch that a developer outside ElevenLabs' platform can act on. Until then it is a strong argument for buying the platform, which is presumably the point.