
Scribe v3 Live Medical vs MAI Transcribe 2 Streaming: Only One Has a Price
- GoogleNEWGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 per 1M tokens · 89 tok/s
- AnthropicNEWAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 per 1M tokens · 97 tok/s
- orcaNEWOrca: OrcaC2 1.02026-11-079 tok/s
- OrcaNEWOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 88 tok/s
- OrcaNEWOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1M tokens · 134 tok/s
- OrcaNEWOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$2.00 / $5.00 per 1M tokens · 87 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 121 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1M tokens · 68 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 90 tok/s
- OrcaOrcaCyber: MiMo V2.6 Flash Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 44 tok/s
- MoonshotMoonshotAI: Kimi K3.1 Pro2026-10-2447Intelligence
- OrcaOrcaCyber: DeepSeek V4.1 Flash Abliterated / Uncensored2026-10-23$0.80 / $1.60 per 1M tokens · 195 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 120 tok/s
- deepseekDeepSeek: DeepSeek V4.1 Flash 10172026-10-1733Intelligence
- DeepSeekDeepSeek: DeepSeek V4.1 Flash2026-10-1740Intelligence
- DeepSeekDeepSeek: DeepSeek V4.1 Flash 09102026-10-17$0.15 / $0.60 per 1M tokens · 187 tok/s
- OrcaOrca: OrcaCyber Zero 1.52026-10-10$3.00 / $7.50 per 1M tokens · 143 tok/s
- openaiOpenAI: GPT-6.1 Sol2026-09-2955Intelligence
- anthropicAnthropic: Claude Sonnet 5.52026-09-2855Intelligence
Line up Scribe v3 Live Medical and MAI-Transcribe-2-Streaming and the first difference is not accuracy, latency or language count. It is that one of them is a product and the other is a feature. MAI-Transcribe-2-Streaming has a model string you pass to an API, a published rate, a preview status, four named Azure regions and two documented integration paths. Scribe v3 Live Medical has a dropdown in an ElevenAgents agent's settings, an early-access gate, and no price of its own because it is folded into a subscription. Microsoft shipped a transcriber; ElevenLabs shipped a capability inside a platform it also sells.
That asymmetry makes the accuracy comparison the least useful part of this page and the procurement comparison the most useful. Still, the numbers deserve stating, because Microsoft's are the aggressive ones: a claimed 2.5% streaming word error rate at 0.13 seconds to a final transcript, against ElevenLabs' 12.8% on clinical mock consultations. Those two figures are five times apart and describe almost nothing in common, which is exactly why this matchup needs care rather than a verdict.
What Microsoft actually shipped
MAI-Transcribe-2-Streaming is the real-time companion to MAI-Transcribe-2, Microsoft AI's in-house batch model. It arrived on 1 October 2026 in public preview, roughly a month after the batch release, and it is served through Azure's speech stack rather than through a separate AI endpoint. Two documented integration paths exist: an OpenAI Realtime-compatible WebSocket, and the Azure Speech SDK, which handles connection management, retries and audio streaming for you. Both return intermediate results that update in place and final results that confirm each segment.
The claims Microsoft attaches to it, labelled as vendor claims because that is what they are: first for accuracy on both final and partial transcripts on Artificial Analysis' streaming evaluation, at 2.5% word error rate with the final transcript ready 0.13 seconds after end of speech and 2.5% on the first partial at 0.12 seconds. The important caveat, which the batch launch shared: the tracker has not plotted this model as a row. Microsoft is claiming a result on a third party's methodology, not citing a third party's published measurement. The board's actual streaming leader at the time of drafting was Grok Voice Transcribe 2.0 at 2.73% and 0.49 seconds.
Everything else about it is ordinary and well documented: 60 languages with automatic continuous language detection, code-switching for blended pairs like Hinglish and Spanglish, noise robustness claims, and availability in Sweden Central, Central US, East US 2 and Southeast Asia. The preview disclaimer is standard Azure language — no service-level agreement, not recommended for production workloads.

What ElevenLabs shipped, and why it is in a different category
Scribe v3 Live Medical sits inside the Scribe v3 family that launched on 9 November 2026, and it is the medical real-time member of it. ElevenLabs' documentation is unambiguous about the boundary: it is available in ElevenAgents only, it is not available through the Speech to Text API, and if you want medical transcription through the API you should use Scribe v2 Medical — the batch model, at $0.20 per hour, which is a normal product with a model ID.
Three consequences follow, and they matter more than any WER figure:
• You cannot price it. It is included in standard ElevenAgents pricing, which bills per conversation minute and bundles the ASR with the agent's LLM, telephony, speech synthesis and tool calls. There is no separable ASR rate, so there is no number to put in a comparison spreadsheet.
• You cannot benchmark it yourself outside the platform. The only published figure is ElevenLabs' own run on PriMock57 — 57 mock primary care consultations — where it scored 12.8% against 15.5% for AssemblyAI 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.
• You cannot send it PHI yet. The docs say plainly: do not send protected health information to Scribe v3 Live Medical until HIPAA eligibility is confirmed, and EU data residency is not available during early access. That is a firmer statement than the marketing page's general HIPAA-eligible line, and it is the one a compliance officer will read.

2.5% against 12.8% is not a five-times gap
It looks like one. It is not, for three reasons, and each of them is a different kind of measurement problem.
The corpora are different. Artificial Analysis' streaming harness measures general speech across many languages and conditions; PriMock57 is 57 simulated GP consultations in which two people discuss symptoms, prescriptions and test results. Clinical consultations are harder for every model — ElevenLabs' own chart shows four competing real-time engines clustered between 15.5% and 17.2% on it, including two models built specifically for medical audio. A 12.8% on that set is a strong result; a 2.5% on a general set is a different achievement.
The conditions are different. Microsoft's number is a claim on a third party's methodology, not a row that third party publishes. ElevenLabs' number is a vendor run on a vendor-chosen public set. Neither is a measurement you can hold against the other without running both on the same audio, and nobody has.
The tasks are different in a way that matters clinically. Microsoft's model is a general transcriber with keyword biasing and a documented clinical-notes use case; it is not fine-tuned on medication names. ElevenLabs' model is a fine-tune of Scribe v3 Live aimed at exactly that vocabulary. On a clean general board the general model should be expected to win. On "he's on metformin 1000 twice a day, empagliflozin 10, and atorvastatin 40" — the sample line on ElevenLabs' medical page — the fine-tune is the one with a reason to be better, and the general model is the one relying on its phrase list.
The latency difference is the one that is real
Microsoft's 0.13 seconds to a final transcript is the figure that should worry a competitor, because it is the kind of number a platform can be built on. Under about 0.2 seconds, caption text reads as simultaneous with speech; a third of a second reads as the transcript chasing the speaker. ElevenLabs does not publish a latency for the medical variant, only that it matches Scribe v3 Live, whose family comparison is a 44% lower semantic WER than Scribe v2 Realtime — a model ElevenLabs documents at roughly 150 ms median time to committed transcript. So the plausible reading is that both are in the same sub-quarter-second neighbourhood and the difference is not perceptible to a person reading along.
Where the difference does show is endpointing, and here ElevenLabs has the more specific story. Its turn prediction uses conversation context to decide whether a caller has finished, so an agent waits through the pauses inside a member ID or a date of birth instead of replying to the first three digits. Microsoft's equivalent is the Voice Live API's advanced end-of-turn detection, which is a property of the agent layer rather than the transcriber. Both solve it. One solves it inside the ASR product and the other solves it one layer up, and if you are assembling your own agent loop, the layer matters.
Cost, side by side, with the holes marked
• Microsoft — $0.54 per audio hour is the introductory rate on MAI-Transcribe-2-Streaming, with no standard price disclosed; Microsoft has run this structure before, and the batch model's early-bird pricing was explicitly time-limited.
• ElevenLabs — no ASR rate exists for Scribe v3 Live Medical. The family anchors on the API are $0.22 per hour for Scribe v3 Live and $0.08 per hour for realtime keyterm prompting, and the API medical product, Scribe v2 Medical, is $0.20 per hour on the batch endpoint.
• At scale — 10,000 minutes of streamed audio is roughly $90 on Microsoft's introductory rate and unquantifiable on ElevenLabs' platform pricing, where the same ten thousand minutes also bills LLM tokens, telephony and voice.
• Custom vocabulary — Microsoft's phrase list is a request parameter with no published surcharge; ElevenLabs includes up to 10,000 custom terms at no extra cost in ElevenAgents, and bills $0.08 per hour for realtime keyterms on the API models.
That last pair is worth pausing on, because 10,000 terms is not a marketing round number — it is roughly the size of a mid-size health system's formulary plus its provider roster. If your clinical vocabulary problem is large, the term budget is a real feature and Microsoft's phrase list, documented as hints rather than forced output, is a different mechanism with different failure modes.
Where the reasoning layer fits
Neither of these models decides anything. A clinical voice agent takes the transcript and does something with it — checks a coverage rule, drafts a note, decides whether to route to a nurse — and that step is a separate model call with its own latency budget. In a real-time pipeline, the reasoning model's time-to-first-token is often a bigger contributor to perceived delay than the transcriber's 0.13 or 0.15 seconds.
That leg is where a router earns its place. OrcaRouter serves 215 models behind one OpenAI-compatible API at 0% markup, provider list prices passed through unchanged, with automatic failover across providers — so the model reasoning over a transcript can be swapped for a faster one, or failed over mid-call, without touching the Azure or ElevenLabs integration. We do not serve Scribe v3 Live Medical or MAI-Transcribe-2-Streaming; the transcription leg stays wherever you bought it.

The verdict, stated as a procurement decision
Choose MAI-Transcribe-2-Streaming if you need a transcriber you can call, price and put in a rate card. It is the only one of the two that is a model rather than a platform feature, it is documented down to the region and the SDK, and its claimed accuracy-latency pair is the strongest published position in streaming ASR right now — with the caveat that the claim is Microsoft's and the board has not plotted it. Build on the preview with the usual caution: no SLA, and an introductory price that will move.
Choose Scribe v3 Live Medical if you are buying an agent, not a transcriber — if the phone system, the turn-taking, the voice and the guardrails are all going to be ElevenLabs' anyway, then a clinical fine-tune inside that stack is worth more than a better general WER you would have to integrate yourself. Test it on your own recordings first, and do not put PHI through it until the documentation stops warning you not to.
If you are genuinely undecided, the tiebreaker is not on either vendor's site: run twenty of your own calls through both, count the errors that would have changed a clinical decision rather than the errors overall, and compare that count to what each option costs you per month including the parts you would otherwise have to build. On that test the 2.5% and the 12.8% will both be wrong, and the ranking may well flip.
