A transcript that turns "metoprolol" into "metronidazole" isn't a typo. It's a beta blocker becoming an antibiotic, and it's sitting in a note that a downstream summarizer will happily treat as fact. That single class of error, the missed or mangled clinical entity, is what separates a medical transcription API from a general-purpose one. It's also the thing most evaluation processes measure last.
Here's the number we'd start with. Universal-3.5 Pro with Medical Mode posts a 3.2% Missed Entity Rate on our medical transcription benchmark. Same audio, same entity list, very different amounts of cleanup work landing on whoever reviews the note.
This post is about how to pick a medical transcription API on evidence rather than a feature grid. It covers which metric actually predicts clinical usability, how five medical speech-to-text APIs compare on model, activation, entity accuracy, price and languages, how to turn Medical Mode on in one parameter, and how PHI and BAA coverage fit into the decision. If you're comparing two specific vendors, we also have a head-to-head on AssemblyAI vs Deepgram for medical transcription.
What a medical transcription API has to do that a general one doesn't
General speech-to-text is optimized for the average word. Clinical audio punishes that. The words that carry the most consequence, such as drug names, dosages, procedures and anatomical terms, are exactly the words that are rare in general training data, phonetically crowded, and often spoken fast by someone who says them forty times a day.
So a medical transcription API has four jobs beyond "produce readable text":
- Get the entities right. Drugs, conditions, procedures, dosages. A 2% overall word error rate means nothing if the 2% is concentrated in medication names.
- Know who said what. Clinician instructions and patient-reported symptoms are different data. A transcript that merges them is a liability.
- Handle real room audio. Exam rooms have doors, sinks, keyboards, and a patient sitting four feet from the mic while a family member talks over them.
- Handle PHI properly. Redaction, a signed BAA, and a clear answer on where audio lives.
Everything below is organized around those four.
Missed entity rate is the metric to evaluate on
Word error rate is the default ASR benchmark, and it's the wrong lens for healthcare. WER treats "the" and "hydralazine" as equally weighted tokens. Missed Entity Rate doesn't. It measures how often the model fails to capture a clinical entity correctly, which is the failure mode that actually reaches a chart. We go deeper on why in MER vs WER for medical transcription.
On our medical benchmark, Universal-3.5 Pro with Medical Mode records a 3.2% MER. On the same audio, Speechmatics Enhanced Medical records 3.6%, Deepgram Nova-3 Medical 4.7%, AWS Transcribe Medical 8.7% and Google Medical Conversation 24.4%. Against the base model without Medical Mode, Medical Mode means roughly 20% fewer missed medical entities and about 87% fewer medical entity errors. The add-on isn't a rounding-error improvement; it's a different error profile.
When you run your own eval (and you should), do it on your audio, not on clean read speech. Pull 20 to 50 real encounters, build an entity list from the notes your clinicians actually wrote, and score entity capture rather than overall WER. That's a two-day exercise that will tell you more than any vendor deck.
What's new in Medical Mode (November 2026)
Medical Mode now runs on an updated correction model that improves medical-term recall. It's the same domain: "medical-v1" parameter, so there's no code change. If you already send medical-v1, your transcripts pick up the update automatically.
The rest of the picture:
- Where it runs. Pre-recorded audio on AssemblyAI-STT-3.7 and Universal-2, and streaming audio on AssemblyAI-STT-3.7-Realtime, AssemblyAI-STT-3.7-Realtime-Agent and Universal-3.6 Pro Realtime. It isn't available on the Dictation API or the Sync API.
- Languages. English, Spanish, French and German, for both pre-recorded and streaming.
- Drug-name misses. 30% fewer in batch and 16% fewer in streaming (English, pooled), with word error rate unchanged, per AssemblyAI's Medical Mode research.
- Medical-term misses. Up to 13% fewer missed medical terms than the base model, with no rise in wrongly inserted terms. On Universal-2, Medical Mode cuts missed medical terms by 27%.
- Price. A $0.15/hr add-on on top of the base model rate.
- Vet Mode. Veterinary audio now has its own mode. Vet Mode is in beta for pre-recorded and streaming audio, with
domain: "veterinary-v1".
Medical transcription API comparison
Here's how five medical speech-to-text APIs compare on the specs that decide an integration. Missed entity rates come from the same benchmark, on the same audio. Other vendors' pricing varies by plan, contract and region, so check their pricing pages before you model costs.
| Provider and model | How you turn on clinical accuracy | Missed entity rate | List price | Clinical languages |
|---|---|---|---|---|
| AssemblyAI Medical Mode on AssemblyAI-STT-3.7 (pre-recorded) and AssemblyAI-STT-3.7-Realtime (streaming) | domain: "medical-v1" on the same model, pre-recorded and streaming |
3.2% (measured on Universal-3.5 Pro with Medical Mode) | Base rate ($0.21/hr pre-recorded, $0.30/hr streaming) plus $0.15/hr Medical Mode | English, Spanish, German, French |
| Speechmatics Enhanced Medical | Medical domain: domain=medical on the Enhanced tier, pre-recorded and streaming |
3.6% | Varies by plan | English, Spanish, French, German |
| Deepgram Nova-3 Medical | Separate model: model=nova-3-medical |
4.7% | Varies by plan | English variants, plus a 10-language multilingual mode |
| AWS Transcribe Medical | Separate API: medical transcription jobs and medical streams | 8.7% | $0.075/min in AWS's pricing example, as of November 2026 | US English only |
| Google Medical Conversation | Separate model: medical_conversation |
24.4% | Varies by plan | US English only |
The short read: AWS and Google give you a medical endpoint inside a cloud you may already be in, limited to US English and with a wide gap on entity accuracy. Deepgram and Google route clinical audio to a separate medical model, so you run two models if you also transcribe non-clinical audio. Speechmatics is the closest on missed entity rate. AssemblyAI's argument is that you shouldn't have to switch models to get clinical accuracy: one parameter on the flagship model gets you there, on both pre-recorded and streaming audio.
Drug names are a separate measurement. AssemblyAI's Medical Mode research compared six systems in batch, averaged over 9 sets in 4 languages, with no custom vocabulary for any vendor. AssemblyAI with Medical Mode had the lowest average WER, medical-term miss rate and drug-name miss rate of the six systems tested in batch. These numbers come from a different evaluation than the missed entity rates in the table above. Average drug-name miss rate in batch:
- AssemblyAI + Medical Mode: 10.7%
- Speechmatics (Enhanced, medical): 11.1%
- ElevenLabs (Scribe v2 Medical): 11.1%
- Google Gemini 3.5 Transcribe: 13.4%
- Soniox (stt-async-v6): 16.7%
- Deepgram (nova-3-medical): 26.6%
Sign up free, add domain: "medical-v1" to your request, and score entity capture against whatever you're running now. No commitment, no sales call.
How Medical Mode works in practice
Medical Mode is a parameter, not a different model. You keep AssemblyAI-STT-3.7 and everything it does, including diarization, formatting and code-switching. Medical Mode adds a correction pass on top.
Here's what that pass is. After the speech model produces the transcript, an LLM reviews it and corrects medical terms that came through misspelled or garbled, mostly medication names and, to a lesser extent, disease names. It follows clinical writing conventions too, such as writing a generic drug name with the brand in parentheses.
That design also sets its limits. Medical Mode corrects what the speech model got close to. It can't recover a term the model never heard, and it won't add information that wasn't spoken. For rare terms specific to your clinic, pair it with keyterms or contextual prompting (below).
What Medical Mode corrects (and what it doesn't)
Medical Mode's gains are concentrated on medication names and, to a lesser extent, disease names. In AssemblyAI's Medical Mode research on 84 hours of English clinical audio across 4 evaluation sets, here's how many misses it removed on net in batch, by category:
- Drug names: +55 of 182 (30%)
- Diseases: +16 of 180 (9%)
- Symptoms: +6 of 328
- Procedures: +1 of 102
- Anatomy: 0
- Dosages: 0
- Tests: −2
Anatomy, dosages, tests and procedures are mostly numbers, units or everyday words, and Medical Mode doesn't measurably help them. If dosage accuracy is your main risk, validate it separately. Word error rate moves by 0.1 point or less on every English set, so the corrections leave the rest of the transcript alone. The gain is larger in batch than in streaming, because streaming corrects turn by turn with less context, under a time budget.
Spanish shows the same pattern. On 4 hours of clinical conversations, the drug-name miss rate fell from 17.6% to 6.7% in batch (−62%) and from 16.7% to 8.6% in streaming (−49%).
These are drug-name and category miss counts, a different measure from the 3.2% missed entity rate above.
Pre-recorded audio
Pre-recorded transcription takes speech_models as a list:
from assemblyai.prerecorded.v2 import Transcriber, TranscriptionConfig
config = TranscriptionConfig(
speech_models=["assemblyai-stt-3-7", "universal-2"],
domain="medical-v1",
speaker_labels=True,
)
transcriber = Transcriber(api_key="<YOUR_API_KEY>")
transcript = transcriber.transcribe(
"https://example.com/encounter-audio.wav",
config,
)
for utterance in transcript.utterances:
print(f"Speaker {utterance.speaker}: {utterance.text}")Live audio
Streaming uses speech_model, singular, and the same domain parameter. This example streams a 16 kHz mono WAV file in 100 ms chunks and prints each finished turn with its speaker label:
import time
import wave
from assemblyai.streaming.v3 import (
StreamingClient,
StreamingClientOptions,
StreamingEvents,
StreamingParameters,
TurnEvent,
)
def on_turn(client, event: TurnEvent):
if event.end_of_turn:
print(f"[{event.speaker_label}] {event.transcript}")
def audio_chunks(path, chunk_ms=100):
with wave.open(path, "rb") as wav:
frames_per_chunk = int(wav.getframerate() * chunk_ms / 1000)
while True:
chunk = wav.readframes(frames_per_chunk)
if not chunk:
break
yield chunk
time.sleep(chunk_ms / 1000) # pace the file like live audio
client = StreamingClient(StreamingClientOptions(api_key="<YOUR_API_KEY>"))
client.on(StreamingEvents.Turn, on_turn)
client.connect(
StreamingParameters(
sample_rate=16000,
speech_model="assemblyai-stt-3-7-realtime",
domain="medical-v1",
mode="balanced",
speaker_labels=True,
)
)
try:
client.stream(audio_chunks("encounter-16khz-mono.wav"))
finally:
client.disconnect(terminate=True)AssemblyAI-STT-3.7-Realtime is the streaming model here, built for ambient transcription with realtime speaker labels. Its turn detection is content-aware: a finished sentence ends the turn and a turn that ends in an entity is held, which keeps the live record segmented the way the conversation actually went.
Contextual prompting
This is the feature we'd point at first if you're building anything longitudinal. You can give the model context before it transcribes, and in a clinical setting the most useful context is the patient's own prior-visit note. In internal healthcare testing, doing exactly that cut missed medical terms by 31%.
Think about why. A patient on lisinopril and metformin who was seen for the same complaint six weeks ago has a vocabulary. Handing the model that vocabulary as free text, rather than a curated keyword list you have to maintain, resolves most of the phonetic ambiguity before it becomes an error. On streaming, contextual prompting with detailed context cut medical-term entity errors by 43% on 20,000 real voice-agent calls.
Medical Mode and prompting do different jobs. Prompting shapes what the model hears; Medical Mode corrects what it wrote. Together they beat either alone.
AssemblyAI vs Amazon Transcribe Medical
This is the comparison healthcare teams on AWS ask about most, so here it is directly.
- Entity accuracy. On the same benchmark audio, AWS Transcribe Medical misses 8.7% of medical entities. Universal-3.5 Pro with Medical Mode misses 3.2%.
- Languages. Transcribe Medical supports US English only. Medical Mode supports English, Spanish, German and French.
- Integration. Transcribe Medical is a separate API from standard Amazon Transcribe. Medical Mode is one parameter on the same AssemblyAI request you'd use for any audio, pre-recorded or streaming.
If you're already deep in AWS, Transcribe Medical is the path of least resistance. Just measure the MER on your own encounters before you commit. Our roundup of Amazon Transcribe alternatives covers the wider field.
Doctor, patient, and the third person in the room
Diarization is where a lot of clinical transcription quietly falls apart. The failure usually isn't a long monologue attributed to the wrong person. It's the short turns. "Mm-hm." "How long?" "Since Tuesday." Those two-word exchanges carry the clinical content, and models tuned for diarization error rate tend to smooth right over them.
AssemblyAI-STT-3.7 runs on a new speaker diarization model. Compared with Universal-3.5 Pro at launch, it cuts cpWER, the measure of getting the right words attributed to the right speaker, by 20% (research). It holds up on short turns and overlapped speech, the two things exam-room audio is made of. In streaming, diarization supports revision as more audio arrives, for up to 10 speakers.
For far-field ambient capture, match your setup to your hardware. Voice Focus (voice_focus, with near-field and far-field settings) runs on AssemblyAI-STT-3.7-Realtime-Agent and Universal-3.6 Pro Realtime, not on AssemblyAI-STT-3.7-Realtime. A phone on a desk three feet from the clinician and a headset mic are not the same acoustic problem.
Multilingual clinical audio
Two separate facts here, and they get conflated constantly.
First: AssemblyAI-STT-3.7 code-switches natively across 32 languages, with no configuration. A patient who moves between English and Spanish mid-sentence, the normal case in a lot of clinics, doesn't break the transcript or require you to detect and route language up front.
Second: Medical Mode itself supports English, Spanish, German and French, for both pre-recorded and streaming audio. So the clinical correction pass covers four languages while the underlying model handles considerably more general speech. If your patient population needs clinical entity accuracy in a fifth language, that's a roadmap conversation, not a config flag.
The fastest comparison is an audible one. Drop a clinical recording into the playground with Medical Mode on and read the entities yourself.
PHI, BAA, and where the audio lives
Compliance shouldn't be the reason you pick a vendor, but it will absolutely be the reason procurement blocks one.
AssemblyAI is considered a business associate under HIPAA and offers a standard Business Associate Addendum (BAA) for customers processing PHI. Paid customers can review and sign it from the dashboard, and the BAA FAQ covers the details. AssemblyAI is also SOC 2 Type 2, ISO 27001:2022 and PCI DSS v4.0 certified.
On the data side, PHI redaction runs across both audio and transcripts. Identifiers are removed from the text and bleeped in the media file, so you're not left with a clean transcript sitting next to an unredacted recording. For teams with residency requirements, there's EU data residency, and a self-hosted deployment option when the audio can't leave your infrastructure at all.
The practical sequencing advice: start the BAA conversation in parallel with your technical eval, not after it. It's rarely the blocker people expect, but it's always the one that adds weeks if you serialize it.
Voice agents and the healthcare front desk
Transcription is one half of the healthcare voice story. The other is the phone line: appointment scheduling, refill requests, intake and triage routing. There you need speech-to-text, an LLM and text-to-speech stitched together with latency low enough that the caller doesn't start talking over the bot.
AssemblyAI's voice agent solutions cover both speech ends of that stack: AssemblyAI-STT-3.7-Realtime-Agent for speech-to-text and AssemblyAI Text-to-Speech for voice output, orchestrated with LiveKit, Pipecat or your own loop. For healthcare specifically, that means the entity accuracy that matters for a refill request ("Is that amoxicillin or amoxicillin-clavulanate?") comes from the same model family you'd use for clinical documentation.
Our partners have been vocal about the context handling in particular:
"We're excited to make AssemblyAI's Universal-3.5 Pro available on LiveKit Inference. What really stands out is their pace of innovation with Context Carryover — it intelligently applies conversation context to improve transcription accuracy in a way most speech models don't, removing the need for users to predefine key terms." — David Zhao, Co-founder at LiveKit
That last clause is the one healthcare teams should notice. Maintaining a keyword list per specialty, per formulary and per clinic is a real operational cost. Not having to is a feature.
How to actually decide
A decision framework that fits on a napkin:
- Already deep in AWS or GCP with modest accuracy requirements? Their medical endpoints are the path of least resistance. Just measure the MER before you commit.
- Building an ambient scribe? Entity accuracy and diarization quality dominate everything else. See our guide to choosing a speech-to-text API for ambient scribes.
- Replacing a general model that's failing on drug names? Medical Mode is one parameter. That's a one-afternoon test.
- Hard data residency constraints? Start with self-hosted and EU residency options and narrow from there.
- Comparing across the whole field? Our roundup of the best medical speech-to-text options goes provider by provider.
Ambient scribe teams are already building on this stack. Here's how Commure describes it:
"We've integrated the newest models from AssemblyAI for pre-recorded audio ASR in our ambient product, and it's been excellent. We're now exploring Universal-3.5 Pro for async and realtime speech-to-text capabilities for new use cases. What's been just as important is the reliability of the platform itself—both technically and in terms of partnership." — Gautam Pradeep, Tech Lead, Commure
Where this goes next
The interesting shift isn't accuracy. Accuracy is a solved-enough problem that the top of the benchmark is getting crowded. It's that context is becoming the primary lever. Contextual prompting means the model's performance on a given encounter now depends on what you tell it beforehand, which turns transcription from a stateless API call into something closer to a personalized service. The 31% reduction in missed terms from a prior-visit note is a preview of that: the same model, the same audio, materially better output, because it knew who it was listening to.
Which means the teams that win the next round won't be the ones who picked the best model. They'll be the ones who built the plumbing to feed it the right context at the right moment: patient history, care setting, specialty and prior encounters. Start designing for that now, even if you're only using it for a keyword hint today.
Our team will help you design the eval, pick the right entity list, and run the comparison properly, including the migration estimate if you decide to switch.
Frequently Asked Questions (FAQs)
What is the best medical transcription API?
The best medical transcription API is the one with the lowest missed entity rate on your own clinical audio. On AssemblyAI's medical benchmark, Universal-3.5 Pro with Medical Mode posts a 3.2% Missed Entity Rate. On the same audio, Speechmatics Enhanced Medical records 3.6%, Deepgram Nova-3 Medical 4.7%, AWS Transcribe Medical 8.7% and Google Medical Conversation 24.4%. Run 20 to 50 of your own encounters through a shortlist and score entity capture, not overall word error rate.
How accurate is AssemblyAI Medical Mode compared to other providers?
Universal-3.5 Pro with Medical Mode has a 3.2% Missed Entity Rate on AssemblyAI's medical benchmark, which also covers Speechmatics, Deepgram, AWS and Google on the same audio. Compared with the base model without Medical Mode, it means roughly 20% fewer missed medical entities and about 87% fewer medical entity errors.
AssemblyAI vs Amazon Transcribe Medical: which is more accurate?
On the same benchmark audio, Universal-3.5 Pro with Medical Mode misses 3.2% of medical entities and AWS Transcribe Medical misses 8.7%. Transcribe Medical supports US English only, while Medical Mode supports English, Spanish, German and French on both pre-recorded and streaming audio.
What changed in the updated Medical Mode?
Medical Mode now uses an updated correction model that improves medical-term recall. It uses the same domain: "medical-v1" parameter, so existing integrations get the update with no code change. It runs on pre-recorded and streaming audio in English, Spanish, German and French. On English clinical audio, it cut drug-name misses by 30% in batch and 16% in streaming, with word error rate unchanged. The gains are concentrated on medication names and, to a lesser extent, disease names.
Does AssemblyAI redact PHI from medical transcripts?
Yes. PHI redaction runs across both the transcript text and the audio itself, so identifiers are removed from the text and obscured in the media file. That matters because a redacted transcript paired with an unredacted recording isn't actually redacted.
Does AssemblyAI sign a BAA?
Yes. AssemblyAI is considered a business associate under HIPAA and offers a standard Business Associate Addendum (BAA) for customers processing PHI. Paid customers can review and sign it self-serve from the dashboard, at no additional cost.