Medical transcription in Spanish, German, and French: multilingual clinical accuracy

Multilingual medical transcription is the transcription of clinical audio in which speakers move between languages, often inside a single sentence. AssemblyAI handles it with Medical Mode (domain: "medical-v1") on Universal-3.5 Pro, which is benchmarked in English, Spanish, German, and French and posts a 3.2% Missed Entity Rate—the lowest across benchmarked providers. The underlying model code-switches natively across 18 languages, so there is no language toggle and no detect-then-route pipeline to maintain.

Most multilingual transcription stories are about coverage. How many languages does the model support? Forty? A hundred?

That's the wrong question for medicine.

In a real clinic, the hard problem isn't supporting French. It's the moment a Spanish-speaking patient in a US clinic says "me duele el pecho, like a pressure" in one breath, and your transcript has to get both halves right—the Spanish symptom and the English qualifier—without anyone touching a language setting. Coverage counts languages. Clinical accuracy counts the words inside a single messy sentence. Those are not the same achievement.

Medical Mode is benchmarked across four languages—English, Spanish, German, and French—for both pre-recorded and streaming audio. But the part I want to spend this post on is the part that actually breaks competing systems: what happens at the seam between languages.

Why multilingual medical transcription is genuinely hard

Let's be honest about the difficulty, because it's easy to wave at "multilingual support" as if it were a switch you flip.

First, medical vocabulary is treacherous across languages precisely because so much of it shares Latin and Greek roots. "Hypertension," "hipertensión," "Hypertonie," and "hypertension" look and sound like cousins—and that similarity is a trap, not a help. A model has to map each one to the right clinical entity in the right language, not blur them into an average. Near-identical isn't identical, and in a medication or diagnosis field, close is wrong.

Second, accents. A francophone clinician in Montreal, a Swabian physician, and a Madrid pharmacist don't pronounce their own languages the way a textbook does, let alone the way a model's training distribution assumes.

Third—and this is the one that quietly wrecks deployments—real clinical encounters are mixed-language. US Hispanic care, the DACH region with its English-heavy medical training, francophone clinics with English drug brand names. People code-switch. They start a sentence in one language and finish it in another, drop an English drug name into a Spanish sentence, or answer a German question in English because that's how they learned the term.

A pipeline built on "detect the language, then transcribe in that language" has no good answer here. By the time it's committed to Spanish, the English half of the sentence is already a casualty.

And notice how that failure presents. It doesn't throw an error. It doesn't flag low confidence. The detector picks Spanish, the Spanish model does its honest best with an English drug name, and out comes a phonetic approximation that looks like a word. A clinician skimming the note sees something plausible in the medication field. That's the worst possible failure mode in clinical documentation: wrong, confident, and quiet.

Which is why "what's your accuracy in Spanish?" is the wrong diligence question. The number you actually want is accuracy on the sentences where the language changes, because those are the sentences that produce silent errors, and they're systematically underrepresented in every vendor's benchmark suite.

What is intra-utterance code-switching in clinical audio?

So here's where it gets interesting.

Universal-3.5 Pro handles native intra-utterance code-switching across all 18 of its languages: English, Spanish, French, German, Italian, Portuguese, Arabic, Danish, Dutch, Finnish, Hebrew, Hindi, Japanese, Mandarin, Norwegian, Swedish, Turkish, and Vietnamese.

Intra-utterance is the operative phrase. Not "we detect a language per file," and not even "per sentence." Within a single utterance, the model follows a speaker as they move between languages and keeps transcribing accurately the entire way through—because code-switching is built into the model rather than stitched on afterward.

For live audio, Universal-3.5 Pro Realtime carries the same native code-switching into streaming, so a clinical encounter that mixes languages mid-sentence holds the thread in real time too. Medical Mode itself is benchmarked across the four launch languages, but the code-switching the model relies on is native across the whole set.

No language toggle. No separate pipeline per language. No detect-then-route logic you have to build and maintain.

Here's what that's worth in measured terms. On a five-language code-switching benchmark (normalized WER, lower is better):

That gap matters more than a language count does. In AssemblyAI's 2026 Voice Agent Report—455 responses fielded across Q4 2025 and Q1 2026—62% of builders said accent and dialect support was a priority when choosing speech infrastructure, and 27.5% reported accent or dialect difficulty as a live problem in what they had already shipped. Code-switching is the harder version of that same problem.

Look at the spread rather than the top line. The gap between first and second is meaningful; the gap between first and fourth is the difference between a transcript a clinician can sign and one they have to rewrite. Code-switched audio is where general-purpose models fall apart, and it falls apart quietly—the transcript still reads fluently, it just says the wrong thing.

That means the Spanglish sentence—"me duele el pecho, like a pressure"—comes through intact, both halves, because the model was never forced to pick a lane. The clinician who slips into English for the drug name and back into French for the symptom doesn't break the transcript. The patient who answers in their first language mid-question doesn't either.

The practical upshot for product teams: you build one integration. Not one per market.

How accurate is medical transcription in Spanish, German, and French?

Code-switching is the headline, but it only matters if the underlying transcription is clinically accurate. So let's tie it back to numbers.

Medical Mode posts a 3.2% Missed Entity Rate. That is the lowest Missed Entity Rate across benchmarked providers. Against the base model running on its own, it misses roughly 20% fewer medical entities and produces roughly 87% fewer entity errors.

MER is the right yardstick here because it measures what a clinician actually cares about: how often a clinically meaningful entity—a drug, a dose, a diagnosis—gets dropped or mangled. And critically, Medical Mode is benchmarked across all four launch languages, so that 3.2% isn't an English-only figure dressed up as multilingual. See the full breakdown on the benchmarks page.

Activation is the same single parameter in every language. This sample targets Python SDK 1.0.0:

# pip install "assemblyai>=1.0.0"
import os

from assemblyai.prerecorded.v2 import Transcriber, TranscriptionConfig

config = TranscriptionConfig(
    speech_models=["universal-3-5-pro"],
    domain="medical-v1",
    speaker_labels=True,
)

transcriber = Transcriber(api_key=os.environ["ASSEMBLYAI_API_KEY"])
transcript = transcriber.transcribe(
    "https://example.com/encounters/madrid-intake.mp3",
    config=config,
)

print(transcript.text)

You can let the model auto-detect language, but you don't need a language toggle to get code-switching—it's native. The domain="medical-v1" flag is identical whether the audio is Spanish, German, French, English, or a mix of them in one recording.

This runs the same way for both modes. Pre-recorded audio goes through Universal-3.5 Pro at $0.21/hr, and live audio runs on Universal-3.5 Pro Realtime at $0.45/hr base, with end-of-turn detection landing around 300ms. Medical Mode adds $0.15/hr in either case—so Universal-3.5 Pro plus Medical Mode is $0.36/hr for pre-recorded audio, and on streaming it is a +$0.15/hr add-on to the $0.45/hr base rate. The full table is on the pricing page.

Teams whose whole business depends on non-English accuracy feel this the most directly. Ebru Yildirim, Founder and CEO of Ollang, put a number on it:

"The 30-40% reduction in speech-to-text errors has significantly improved our production efficiency and client satisfaction. We've achieved industry-leading word error rates for non-English audio, which is critical for serving our enterprise clients."

Note what that reduction buys, because it isn't a nicer transcript. It's fewer transcripts that need a human to fix them, which is the only place multilingual accuracy shows up on a P&L.

If you want the wider context on where this fits in a clinical stack, we cover it in AI medical transcription in healthcare and in the ambient AI scribe explainer.

Region-specific drug names and keyterms prompting

One more wrinkle that matters specifically for multilingual healthcare: drug brand names don't translate, they vary by market.

How do I handle drug brand names that differ by market?

The same molecule ships under different brand names in the US, Germany, France, and Spain. A general medical model can't possibly anticipate every regional formulary. So when you're deploying into the DACH region or francophone care, keyterms prompting lets you hand the model the specific brand names, local terminology, and formulary your clinicians actually use—before it processes a single second of audio. Universal-3.5 Pro accepts up to 1,000 keyterms, which comfortably covers a regional formulary. It's the difference between a model that knows the active ingredient and one that also knows what your pharmacist calls it.

Pair that with speaker diarization for multi-speaker encounters and PHI redaction across both audio and transcripts—$0.08/hr for text, $0.05/hr for audio—and you've got a clinical pipeline that holds up across markets.

On the privacy side: AssemblyAI enables covered entities and their business associates subject to HIPAA to use the AssemblyAI services to process protected health information (PHI). AssemblyAI is considered a business associate under HIPAA, and we offer a standard Business Associate Addendum (BAA) that is required under HIPAA to ensure that AssemblyAI appropriately safeguards PHI—on BAA-eligible infrastructure, with the BAA included. It's pre-signed and self-serve in the dashboard, so you can execute it in minutes without a sales call.

And for European deployments, which is most of the reason you're reading a post about German and French: the same models run on api.eu.assemblyai.com and streaming.eu.assemblyai.com at the same price, with data staying in the EU. That's usually the second question a European clinical buyer asks, right after accuracy.

Which languages route where

Being precise about coverage is worth more to you than a big round number, so here's the honest map.

Eighteen languages get native code-switching on Universal-3.5 Pro async, and those are the ones listed above. Medical Mode's clinical benchmarking covers four of them: English, Spanish, German, and French, on both pre-recorded and streaming audio. Audio in a language Medical Mode doesn't cover returns a warning rather than a charge for the add-on, so you won't quietly pay for something that isn't running.

Read the second row carefully, because it's the one that catches people out. A Dutch or Hindi encounter still gets the flagship model and still gets native code-switching. It just doesn't get the clinical entity tuning, and you don't pay for an add-on you aren't receiving. For a lot of deployments that's an entirely acceptable trade. For a psychiatric practice in Amsterdam it's the thing to raise with us before you build.

If your caseload runs outside those 18, that's a conversation worth having early rather than discovering in a pilot—coverage beyond the flagship line exists, and the right path depends on whether that market's clinical audio is genuinely monolingual or mixed.

That's the useful version of a coverage claim. Not "we support N languages," but "here is exactly what happens to your audio, by language, and here is where the behavior changes." Vendors who lead with the biggest number are usually the ones who can't answer the second question.

The integration you don't have to fork

Here's the insight worth sitting with if you're choosing an approach for a multilingual patient population.

The hidden cost of "supports N languages" architectures isn't accuracy—it's branching. A detect-then-route pipeline forces you to maintain a code path per language, test each one, and handle the seams between them, and those seams are exactly where mixed-language encounters live. Native code-switching collapses that whole tree into one path. You're not just getting a better transcript on Spanglish; you're getting an integration that doesn't fork every time you enter a new market.

And notice which cost that actually is. It isn't a line on your inference bill, it's a line in your test matrix—the one that grows multiplicatively while everyone is looking at the per-hour rate. Collapse the branches and the second market costs roughly what the first one did.

Build one pipeline for the 18 languages the model covers natively, know where the boundary sits, and serve the patient who reaches for a different language mid-sentence. If you want to see the full clinical picture, start with Voice AI for healthcare.

Frequently asked questions

Which languages does Medical Mode support?

Medical Mode is benchmarked across four languages—English, Spanish, German, and French—for both pre-recorded and streaming audio. The underlying Universal-3.5 Pro model transcribes 18 languages with native code-switching, and audio in a language Medical Mode doesn't cover returns a warning rather than a charge for the add-on.

What is intra-utterance code-switching, and why does it matter for clinics?

Intra-utterance code-switching means the model follows a speaker who changes languages inside a single sentence, not just per file or per sentence. In clinics where patients and clinicians mix languages mid-thought, it keeps the transcript intact without a language toggle or a separate pipeline per language.

How accurate is multilingual medical transcription?

Medical Mode posts a 3.2% Missed Entity Rate, the lowest Missed Entity Rate across benchmarked providers, and around 20% fewer missed medical entities than the base model. On a five-language code-switching benchmark, Universal-3.5 Pro averages 7.69 normalized WER against ElevenLabs Scribe v2 at 8.77 and Deepgram Nova-3 Multilingual at 12.22.

What is the best medical transcription software for multilingual clinics?

For mixed-language encounters the deciding factor is native code-switching, not language count—a system that detects one language per file will lose the other half of a Spanglish sentence. AssemblyAI's Medical Mode runs on Universal-3.5 Pro, which code-switches natively across 18 languages, and activates with a single domain: "medical-v1" parameter at $0.15/hr on top of the base model.

How do I handle region-specific drug brand names?

Use keyterms prompting to supply the specific brand names, local terminology, and formulary your clinicians use before transcription runs. Universal-3.5 Pro accepts up to 1,000 keyterms, which is enough to cover a full regional formulary.

Does code-switching require a different setup than single-language transcription?

No. You activate Medical Mode with the single domain: "medical-v1" parameter and code-switching is native—there is no toggle to enable and no separate multilingual pipeline to configure. On streaming, if a session is genuinely monolingual you can pass language_codes as a single-element list, for example ["es"], to steer the model toward one language; leave it off to keep native code-switching.

‍