Title card reading 'Scribe v3 Live Medical vs Gemini 3.5 Transcribe Live' with the subtitle '12.8% and 4.00% are not the same number'. Below it two horizontal bars labelled 'PriMock57 clinical mock consultations' at 12.8% and 'Artificial Analysis general speech' at 4.00%, each bar tagged with the benchmark it belongs to and separated by a dashed vertical divider.
Guides & Insights

Scribe v3 Live Medical vs Gemini 3.5 Transcribe Live: 12.8% and 4.00% Are Not the Same Number

Author

jinhao.song

Date Published

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

Put Scribe v3 Live Medical's 12.8% next to Gemini 3.5 Transcribe Live's 4.00% and the streaming transcription question looks settled in Google's favour by a factor of three. It is not, and the reason is the most useful thing on this page: those two numbers were measured on different audio, by different organisations, for different purposes. The 12.8% is ElevenLabs' word error rate on PriMock57 — 57 recorded mock primary-care consultations, full of drug names, dosages and half-finished sentences. The 4.00% is Artificial Analysis' streaming word error rate for Gemini 3.5 Transcribe Live on its general speech harness, at 0.40 seconds to a final transcript. One is a clinical consultation; the other is a benchmark suite. Comparing them is the same mistake as comparing a car's fuel economy on a track to its fuel economy in city traffic and concluding the engine changed.

Strip the false head-to-head away and a real choice remains, and it is about deployment rather than accuracy. Gemini 3.5 Transcribe Live is a model you call: gemini-3.5-transcribe-live, a WebSocket endpoint with a published per-minute price, a free tier, and a ten-minute session ceiling. Scribe v3 Live Medical is a model you cannot call at all — it exists only as an ASR option inside ElevenAgents, in early access, with no model ID and no rate of its own. One is infrastructure. The other is a feature of an agent platform. Which you want depends on whether you are building a transcriber or buying a phone agent.

What each side actually publishes

The spec contrast, dimension by dimension, as each vendor states it:

• Access — Gemini 3.5 Transcribe Live is a Live API WebSocket endpoint (gemini-3.5-transcribe-live) with a free tier; Scribe v3 Live Medical is selectable only in an ElevenAgents agent's Automatic Speech Recognition settings, and not through the Speech to Text API.

• Price — Gemini 3.5 Transcribe Live is $3.50 per million audio input tokens and $21.00 per million text output tokens, which Google estimates at $0.005 and $0.004 per minute respectively, about $0.009 per minute blended; Scribe v3 Live Medical has no separate price and is included in ElevenAgents pricing.

• Accuracy — Gemini: 4.00% streaming WER at 0.40s to final, per Artificial Analysis; Scribe: 12.8% WER on PriMock57, vendor-run, against 15.5%–17.2% for four competing real-time models.

• Domain adaptation — Gemini: custom vocabulary biasing on the batch sibling, with diarization and word timestamps described on the Gemini 3.5 Transcribe card; Scribe: a clinical fine-tune plus up to 10,000 custom terms at no extra cost.

• Languages — Gemini: 85+ languages with mid-stream switching; Scribe: 90+ languages, medical gains measured in English, French and German.

• Session shape — Gemini: a WebSocket session capped at ten minutes; Scribe: lives for the length of the agent conversation.

• Compliance — Gemini: no medical or HIPAA positioning that we could find on the model documentation; Scribe: HIPAA-eligible with a BAA and Zero Retention Mode, with an explicit warning not to send PHI to the early-access real-time model until its own eligibility is confirmed.

Two of those rows decide most real decisions, and neither is accuracy. The ten-minute cap is architectural: a session has to be stitched for anything longer than a call centre survey, and the seam is your problem. The absence of a price for the ElevenLabs model is the other: you cannot put it on a rate card, because it is inside a subscription that also bills the LLM, the telephony and the voice.

A two-column scoreboard titled 'Scribe v3 Live Medical vs Gemini 3.5 Transcribe Live — the scoreboard', with the rows Access (ElevenAgents only vs Live API WebSocket), Price (bundled, no ASR rate vs about $0.009 per minute), Accuracy (12.8% PriMock57 vs 4.00% AA streaming WER), Adaptation (clinical fine-tune, 10,000 terms vs custom vocabulary biasing), Languages (90+ vs 85+) and Session (length of the agent call vs 10-minute cap), and a footer reading 'ElevenLabs figures vendor-run on PriMock57; Gemini figures per Artificial Analysis.'

Turning the accuracy row into something you can use

There is a legitimate way to compare these two, and it requires running them, which is exactly what the marketing charts exist to save you from. What can be said from the published evidence is narrower:

On clinical conversations specifically, ElevenLabs has a measured number and Google does not. That is not a claim that Gemini 3.5 Transcribe Live is bad at drug names — it is a claim that nobody has published a PriMock57-equivalent figure for it. Google's custom vocabulary mechanism is the tool it offers for exactly this problem, and vocabulary biasing on the batch model is documented; on the Live endpoint the documentation is thinner than on its batch sibling, and the diarization and word-timestamp features described on the shared Gemini 3.5 Transcribe card are not explicitly confirmed per-endpoint.

On general conversational audio, Google has an independently published number and ElevenLabs has a vendor-run one. The 4.00% comes from a tracker with a fixed harness and dozens of models; the 12.8% comes from a single 57-consultation set that ElevenLabs chose and ran. Both are useful. Neither survives being subtracted from the other.

The practical version: if your workload is clinical calls, borrow a PriMock57-style sample of your own recordings — two speakers, a real formulary in the room, at least one caller reading out a member ID — and run both. If it is general-purpose streaming transcription with a bit of jargon, Google's published number is the more auditable starting point and its vocabulary biasing will carry most of the domain load.

The Artificial Analysis Speech to Text leaderboard, showing the heading 'Speech to Text AI Model & Provider Leaderboard' with the subtitle 'Compare word error rate, speed, and pricing across Speech to Text models and providers', the chart label 'AA-WER Index (Non-streaming)', and a grid of model bars with percentage labels.

The clinical difference is in the pauses, not the words

The feature ElevenLabs built that Google's docs do not describe is turn prediction. Scribe v3 Live Medical's turn model uses conversation context rather than silence alone to decide a caller has finished, which is the difference between an agent that answers after "my member ID is 4-4-2" and one that waits for the last digit. ElevenLabs' own examples are exactly the strings that break silence-based endpointing: member IDs, dates of birth, phone numbers, dosages spelled out.

Gemini 3.5 Transcribe Live pushes the same problem one layer up the stack. Its documentation describes it as a bidirectional streaming transcription model; where the turn ends is decided by whatever agent framework sits above it, and Google's voice-agent story routes through the Live API's own turn handling rather than a transcription-specific turn model. That is a perfectly good architecture and it is a different one: you own the endpointing logic, and you will tune it. On a clinical call where a patient pauses mid-symptom to think, that tuning is not cosmetic.

The cost comparison, done honestly

Google's meter is transparent enough to build a number on. Take a clinic running 40 hours of live calls a day, 22 days a month: 52,800 minutes. At the estimated $0.009 per minute blended, that is about $475 a month for the transcription leg, and it is a line item you can forecast, cap and argue about at renewal.

The ElevenLabs side does not produce a comparable figure, and anyone who quotes one is guessing. The nearest published anchor is the API rate card — Scribe v3 Live at $0.22 per hour, keyterm prompting $0.08 per hour — but Scribe v3 Live Medical is not sold that way. It is inside ElevenAgents, priced per minute of conversation, and the conversation includes the LLM tokens, the speech synthesis and the telephony. That bundling is either a discount or a lock-in depending on what you would otherwise have assembled, and it cannot be settled from a pricing page.

One thing that does transfer: if you end up running both — Google for the transcription, an LLM of your choosing for the reasoning — the second leg is where the routing decisions live. OrcaRouter serves 215 models behind a single OpenAI-compatible API at 0% markup, provider list prices passed through unchanged, so the model reading Gemini's transcript can be swapped or failed over without touching the ASR code. We do not serve either transcription model on this page, and nothing above is our price.

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 to pick

Pick Gemini 3.5 Transcribe Live if you are building the product rather than buying one. It is the only one of the two that is a callable model with a published price, a free tier for evaluation, and an endpoint you can put behind your own agent loop. The ten-minute session cap is the constraint to design around, and the absence of a clinical benchmark is the gap to close with your own audio and a custom vocabulary list.

Pick Scribe v3 Live Medical if you have already decided the agent is ElevenLabs' problem — and if your compliance position allows an early-access model, which the documentation currently says it does not, for PHI, until eligibility is confirmed. The clinical fine-tune and the turn model are real advantages in a live patient call, and the PriMock57 gap to the competing real-time models is the strongest published evidence that clinical adaptation helps at all. It simply is not evidence about Gemini, and Gemini is not evidence about it.

The number to watch is not either WER. It is whether Scribe v3 Live Medical ever gets a model ID of its own. The day it does, this comparison becomes a normal one — two callable endpoints, two meters, one benchmark you can run — and the 12.8% becomes something a buyer can actually check.

Compared in this article1

Detected from this article · Benchmarks: Artificial Analysis · updated daily