Title card reading 'Scribe v3 Live Medical vs Whisper Large v3 Turbo' with the subtitle 'one you cannot rent, one you can own'. Below it a single scale: on the left a locked padlock icon labelled 'ElevenAgents only, no model ID', on the right an open box labelled '809M parameters, MIT licence, runs on about 6 GB of VRAM'.
Guides & Insights

Scribe v3 Live Medical vs Whisper Large v3 Turbo: The Model You Can't Rent vs the One You Can Own

Author

jinhao.song

Date Published

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

Whisper Large v3 Turbo is an 809-million-parameter model under the MIT licence that you can download, quantize to roughly 1.6 GB and run on a single modest GPU at about eight times the speed of Whisper Large v3. Scribe v3 Live Medical is a clinical fine-tune from ElevenLabs that you cannot download, cannot call through an API, and cannot obtain at all unless you are running an agent inside ElevenAgents. Putting them in the same sentence is odd until you notice that they occupy the same slot in a real decision: both are what you reach for when a general-purpose transcriber keeps mangling "empagliflozin", and only one of them is available to most teams.

That is the actual comparison. Not accuracy — the two numbers do not meet — but availability, and what the open baseline costs you in engineering to close a gap a vendor will not sell you a closed one for.

The eleven-word version of why this is not a benchmark fight

ElevenLabs' published figure for Scribe v3 Live Medical is 12.8% word error rate on PriMock57, 57 mock primary care consultations, against 15.5% to 17.2% for four competing real-time medical and general models. Whisper Large v3 Turbo's published figures are 2.1% on LibriSpeech test-clean and 4.2% on test-other — read English audiobook speech, a corpus that has been the standard open benchmark for years and is roughly as far from a phone call about a prescription as audio gets.

On LibriSpeech conditions, an 809M open model beats a clinical fine-tune by a wide margin. On a real consultation with a bad headset, both numbers are noise. Neither figure predicts the other, and anyone who lines them up and announces a winner is comparing a clean-room measurement of a batch model to a vendor-run measurement of a streaming one. The honest framing is that the medical fine-tune claims an advantage over other real-time models on clinical audio, and Whisper claims an advantage over other open models on clean read speech, and neither claim touches the other.

What Whisper Large v3 Turbo actually is, and is not

OpenAI's distillation of Whisper Large v3, released under MIT in October 2024: 32 encoder layers and 4 decoder layers, down from 32, which is where the speed comes from. It runs in faster-whisper, whisper.cpp, ONNX and GGUF builds, quantizes small enough for a laptop, covers 99 languages, and costs nothing in licence fees whether you run it on one GPU or a rack.

What it is not, and this is the list that matters for a clinical agent:

• It does not stream. Whisper's architecture is built around a fixed 30-second window, so "real-time" means chunking audio and re-running, paying for the overlap and accepting that the transcript lags the speech by however much you buffer. There is no native partial-result stream and no turn detection.

• It does not diarize. One stream of words, no concept of who said what. For a two-person consultation where the patient and the clinician both name drugs, you bolt on a separate speaker-labelling model or you accept an unlabelled transcript.

• It does not format. Punctuation and casing come out as the audio produced them, disfluencies and all. No self-correction resolution, no sentence casing, no clinical note structure.

• It is not trained for translation, and it has no domain bias mechanism. There is no keyterm list to hand it the formulary, and no prompt to tell it this is a cardiology call. You fine-tune, or you post-process, or you accept the errors.

Scribe v3 Live Medical, by contrast, is built almost entirely out of those four missing things: it streams, it has turn prediction that waits out a member ID, it carries up to 10,000 custom clinical terms at no charge, and it is a fine-tune of a model trained on conversation rather than audiobooks. ElevenLabs is candid about where that stops — isolated drug names with no surrounding context are the hardest case for every model, and the advice is to use keyterms to bias toward what you expect.

The openai/whisper repository page on GitHub, showing 113k stars and 13.7k forks, the README heading 'Robust Speech Recognition via Large-Scale Weak Supervision' and an MIT license badge in the sidebar.

The engineering cost of closing the gap yourself

This is where the comparison becomes a spreadsheet rather than a leaderboard. A team that wants Whisper Large v3 Turbo to behave like a clinical real-time transcriber has to build four things:

• A streaming wrapper — chunked audio, overlapping windows, partial-hypothesis merging, and an endpointing rule of your own, because Whisper gives you none. Expect to spend real time on the seam handling; it is where most homegrown voice agents get their worst latency.

• A vocabulary mechanism — Whisper has none. Either you fine-tune on your own clinical audio, which needs labelled data you probably do not have, or you post-correct against a drug-name list, which is a fuzzy-matching problem that produces its own class of confident error.

• A diarization pass — a second model, a second latency budget, and a merge step that has to align speaker turns to Whisper's word timings.

• Infrastructure — GPUs, autoscaling, a queue for the batch path, and someone on call when the box that transcribes 40 hours a day falls over at 2 a.m.

None of that is prohibitive. All of it is work that a managed clinical model would have absorbed, and the honest question is whether your team's time is better spent there or on the clinical product itself. For a research group or a platform team that already runs inference infrastructure, the answer is often yes — that is why Whisper became the default in the first place. For a five-person digital health startup whose differentiator is the nurse workflow, it is usually no.

The licence is the real advantage, and it is not a small one

Whisper's MIT licence means no vendor, no meter, no per-minute fee that grows with call volume, and no requirement that patient audio leave your network. For a covered entity, that last point is not a nice-to-have. A self-hosted Whisper deployment inside your own VPC is a compliance story your security team already knows how to approve: PHI never crosses a vendor boundary, log retention is yours to set, and there is no Business Associate Agreement to negotiate.

Against that, the ElevenLabs documentation currently says something a compliance officer will read twice: do not send protected health information to Scribe v3 Live Medical until HIPAA eligibility is confirmed for the model, and EU data residency is not available during early access. The medical transcription page's general statement — HIPAA-eligible for Enterprise customers with a BAA and Zero Retention Mode — describes the platform and the GA medical product, Scribe v2 Medical. For the early-access real-time model, the warning is the operative sentence.

So the open baseline wins the compliance comparison outright today, for a reason that has nothing to do with accuracy and everything to do with paperwork that is not finished.

Marginal cost, at three different volumes

Whisper has a hardware cost and no meter. Run it on a GPU you already own and the marginal cost of a transcribed minute approaches electricity; rent one and it is the instance rate divided by throughput, which on modern serving stacks is a small number of cents per audio hour. At a thousand hours a month, a single rented GPU covers the load with headroom.

Scribe v3 Live Medical has no marginal cost to compute because it is not sold by the minute — it is inside an ElevenAgents subscription that bills per conversation minute and includes the LLM, the telephony and the voice. ElevenLabs' API anchors, for the models you can actually meter, are $0.22 per hour for Scribe v3 Live and $0.20 per hour for Scribe v2 Medical, with realtime keyterm prompting at $0.08 per hour.

At ten hours a month, the managed route is obviously cheaper: no GPU to buy, no pipeline to maintain, and the subscription includes the agent. At ten thousand hours a month, self-hosted Whisper is cheaper than any per-minute meter that has ever been published, and no vendor will ever undercut the marginal cost of your own hardware running MIT weights. The crossover, as always, sits where your engineering salary divided by the hours saved lands below the per-minute rate — and for a small clinical team it lands later than the spreadsheet suggests, because the pipeline you build is a pipeline you maintain forever.

A two-column scoreboard titled 'Scribe v3 Live Medical vs Whisper Large v3 Turbo — the scoreboard', with the rows Licence (platform only vs MIT), Access (ElevenAgents only vs download and self-host), Streaming (native, turn prediction vs no, 30-second windows), Diarization (built in vs no), Vocabulary (10,000 terms included vs none, fine-tune or correct) and Cost (bundled, no ASR rate vs your own GPU), and a footer reading 'Scribe figures vendor-run and unavailable outside ElevenAgents; Whisper figures per OpenAI.'

Where a router fits, precisely

Whisper hands you text; something still has to read it and act. In a clinical workflow that is a second model — drafting the note, extracting the medication list, deciding whether the call needs a nurse — and it is the leg that scales with conversation length rather than audio length. OrcaRouter serves 215 models behind one OpenAI-compatible API at 0% markup, provider list prices passed through unchanged, with automatic failover, so that reasoning model can be a small fast one for turn-by-turn replies and a larger one for the end-of-call note, on the same key.

Be clear about the boundaries: we do not host Whisper Large v3 Turbo and we do not host Scribe v3 Live Medical or any ElevenLabs model. Self-hosted Whisper is your GPU; ElevenLabs' medical model is their platform. What we route is the layer above both.

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.

Which one, for which team

Take Whisper Large v3 Turbo if any of the following is true: you need the transcript to stay inside your own perimeter, you are transcribing recorded audio rather than live calls, your volume is high enough that a per-minute meter is the dominant cost, or you have the engineering capacity to build the streaming, diarization and vocabulary layers yourself. On recorded clinical audio — dictated notes reviewed after the visit — Whisper plus a drug-name post-correction pass is a genuinely strong and genuinely cheap answer, and its MIT licence removes a whole category of procurement conversation.

Take Scribe v3 Live Medical if the audio is live, the vocabulary is clinical, and the agent is already going to be ElevenLabs'. The turn prediction and the 10,000-term budget are things you would otherwise build, the 12.8%-on-PriMock57 result is the vendor's own but the direction is consistent with what every clinical ASR vendor reports, and the platform absorbs the infrastructure. Then wait on PHI until the early-access warning comes off the docs page.

And if you are outside ElevenAgents entirely — which is most people reading this — the real choice is between Whisper and a callable medical model like Scribe v2 Medical, at $0.20 per hour with 15% fewer errors than Scribe v2 on clinical test audio. Scribe v3 Live Medical is the better model on paper. It is simply not on offer, and a model you cannot run is not a competitor to one you can.