Search "best medical speech-to-text" and you get three different kinds of answer. A dictation app is software a clinician talks into to fill a note field. An ambient scribe is software that listens to the whole visit and drafts the note. A speech-to-text API is the layer developers build both of those on. Microsoft Dragon Copilot and AssemblyAI-STT-3.7 are both correct answers to the question, and they aren't substitutes for each other. Figuring out which category you're shopping in is the first and most consequential decision.
The second is which accuracy number you believe. Word error rate is the industry default, and it's misleading for clinical audio, because it weights "the" and "hydrochlorothiazide" identically. Missed Entity Rate (MER) is the metric that predicts whether a note needs rework. On the medical benchmark published on the medical transcription use case page, Universal-3.5 Pro with Medical Mode records 3.2% MER, ahead of Speechmatics (3.6%), Deepgram (4.7%), AWS (8.7%) and Google (24.4%).
What follows is a straight comparison: the three ways medical speech-to-text gets deployed, the criteria worth evaluating on, and a provider-by-provider rundown.
Three ways medical speech-to-text gets used
These have different accuracy profiles, different latency requirements and different vendors. Sorting your use case into one of them narrows the field immediately.
Front-end dictation
The clinician talks, text appears, and they edit it in place. The audio is cooperative: deliberate speech, a close mic, a quiet room. Latency matters a lot because the clinician is watching words land. Accuracy requirements are high, but a human is right there to correct errors, which changes the risk calculus.
Back-end transcription
Audio is recorded, submitted and transcribed after the fact: dictated notes, procedure reports, correspondence. Latency is irrelevant. Accuracy matters more than in dictation because nobody is watching the text form, and a batch pipeline gets you the best available model quality.
Ambient scribing
A microphone captures the whole encounter, and the system produces a note from the conversation. This is the hardest of the three: far-field audio, multiple speakers, overlapped speech, unscripted patient language, and an LLM downstream that will confidently propagate any entity error into a clinical assertion. The guide to speech-to-text for ambient scribes goes deeper on this one, and the best ambient AI scribes roundup covers the finished products.
What "best" should mean when you evaluate
Five criteria, roughly in order of how much they should influence your choice:
- Medical entity accuracy. How often do drugs, dosages, conditions and procedures survive intact? Measure with Missed Entity Rate, not WER. The MER vs WER explainer covers why.
- Speaker attribution. Can it separate clinician from patient from caregiver, and does it hold up on two-word turns?
- Real-world audio robustness. Performance on your mic setup and your room noise, not on studio recordings.
- PHI handling and BAA availability. Redaction scope, who signs a Business Associate Addendum (BAA) and on which plan, residency and deployment options.
- Cost structure. Not the list price: the commit terms and the effective rate at your volume.
Notice what isn't on that list: number of supported languages, number of features and vendor size. Those come up constantly in RFPs and predict little about whether the transcripts will be usable.
Medical speech-to-text comparison
MER figures below come from a single published table on the medical transcription use case page, so they're comparable with each other. Options not in that table are marked as such rather than estimated.
| Option | Type | Medical option | Medical MER (use-case table) | Streaming | BAA | Best fit |
|---|---|---|---|---|---|---|
| AssemblyAI (3.2% measured on Universal-3.5 Pro; Medical Mode now runs on AssemblyAI-STT-3.7 / AssemblyAI-STT-3.7-Realtime) | API | Medical Mode, domain: "medical-v1" |
3.2% | Yes, same parameter | BAA available; paid accounts sign self-serve | Building a scribe or clinical voice product |
| Speechmatics | API | Enhanced Medical | 3.6% | Check vendor docs | Ask the vendor | Teams evaluating dedicated medical models |
| Deepgram | API | Nova-3 Medical (separate model) | 4.7% | Yes | Enterprise customers | Teams already on Deepgram's stack |
| Amazon Transcribe Medical | API | Separate medical endpoint | 8.7% | Yes | Accepted in AWS Artifact | Teams standardized on AWS |
| Google Cloud Speech-to-Text | API | Medical conversation and dictation models | 24.4% | Check vendor docs | Google Cloud BAA | Teams standardized on GCP |
| Google MedASR | Open-weight model | Medical dictation (English) | Not in table | You host it | You host it | Self-hosted dictation experiments |
| ElevenLabs Scribe v2 | API | Scribe v2 Medical | Not in table | Check vendor docs | Through sales | General transcription |
| Microsoft Azure AI Speech | API | Custom speech and vocabulary | Not in table | Check vendor docs | Microsoft BAA for Azure services | Microsoft-centric health IT |
| NVIDIA Riva | Self-hosted toolkit | You train it | Not in table | Check vendor docs | You host it | Air-gapped deployments |
| Microsoft Dragon Copilot, Abridge | Finished product | Built in | Not comparable | Built in | Vendor contract | Buying, not building |
| AssemblyAI Vet Mode (beta) | API | domain: "veterinary-v1" |
No public metric yet | Yes, same parameter | Not applicable (veterinary) | Veterinary scribes and records |
Sign up free and transcribe your own clinical audio with Medical Mode on. Twenty minutes gets you a better answer than twenty pages of comparison.
Medical benchmark: drug names and medical terms
The MER figures above come from the use-case table. A separate benchmark in AssemblyAI's Medical Mode research scores two narrower measures, medical-term miss rate and drug-name miss rate, alongside WER. Methodology: averages over 9 evaluation sets in 4 languages, with no custom vocabulary for any vendor.
In batch, AssemblyAI with Medical Mode had the lowest average WER, medical-term miss rate and drug-name miss rate of the six systems tested, and the best mean rank across the nine sets on all three measures, but it doesn't lead every row. It was best or tied on drug names in 6 of 9 sets.
| System | WER | Medical-term miss | Drug-name miss | |---|---|---|---| | AssemblyAI + Medical Mode | 9.5% | 5.2% | 10.7% | | Speechmatics (Enhanced, medical) | 9.7% | 6.1% | 11.1% | | Soniox (stt-async-v6) | 10.2% | 6.1% | 16.7% | | Google Gemini 3.5 Transcribe | 10.3% | 6.4% | 13.4% | | Deepgram (nova-3-medical) | 11.2% | 9.7% | 26.6% | | ElevenLabs (Scribe v2 Medical) | 11.3% | 5.3% | 11.1% |
Streaming is a trade-off between error and latency. Speechmatics recorded lower streaming WER (9.6%) and drug-name miss (10.9%) than AssemblyAI with Medical Mode, at 4.71 s p90 time-to-final. Among systems that finalize words within about a second, AssemblyAI with Medical Mode had the lowest error rates. Medical Mode adds about 0.65 s at p90.
| System | p90 time-to-final | WER | Term miss | Drug miss | |---|---|---|---|---| | AssemblyAI + Medical Mode | 1.28 s | 10.4% | 6.3% | 13.0% | | AssemblyAI without Medical Mode | 0.63 s | 10.4% | 6.7% | 15.7% | | ElevenLabs Scribe v2 Realtime | 0.80 s | 13.5% | 9.5% | 19.6% | | Deepgram nova-3-medical | 0.87 s | 13.2% | 12.4% | 29.4% | | Soniox stt-rt-v6 | 5.18 s | 10.2% | 6.4% | 16.4% | | Speechmatics Enhanced medical | 4.71 s | 9.6% | 6.1% | 10.9% |
These figures measure a different thing from the 3.2% MER on the use-case table, so read the two sets separately rather than side by side. Medical Mode's gains are concentrated on medication names and, to a lesser extent, disease names.
The options, one by one
1. AssemblyAI: AssemblyAI-STT-3.7 with Medical Mode
Clinical accuracy is a parameter here, not a separate product. Medical Mode is domain: "medical-v1" on AssemblyAI-STT-3.7 for pre-recorded audio and on AssemblyAI-STT-3.7-Realtime for streaming: same models, same endpoints, same features. On Universal-3.5 Pro it recorded 3.2% MER, with roughly 20% fewer missed medical entities and about 87% fewer medical entity errors than the base model.
How it works: Medical Mode is an LLM post-processing pass. The speech model transcribes first, then a correction step uses the surrounding clinical context to fix medical terms the model heard but spelled wrong. Medical Mode now runs on an updated correction model that improves medical-term recall, with the same domain: "medical-v1" parameter and no code change. Gains come mainly on medication names and, to a lesser extent, disease names; it doesn't measurably change dosages, anatomy, tests or procedures, which are numbers, units or everyday words. Word error rate moves by 0.1 point or less on every English set, so the rest of the transcript is left alone.
What else differentiates it: a new speaker diarization model with the lowest diarization error rate of the pre-recorded APIs we measured (10.7%, ahead of ElevenLabs Scribe v3 at 12.0%); native code-switching across 32 languages on pre-recorded audio; and contextual prompting, where detailed context cut medical-term entity errors by 43% on a public benchmark of 20,000 calls. Medical Mode is a $0.15/hr add-on.
Limits to know: Medical Mode corrects misspellings, so a wrong-but-valid word can slip through. It covers four languages (English, Spanish, German and French). It isn't available on the Dictation API or the Sync API. In streaming, Medical Mode adds about 0.65 s at p90 time-to-final.
2. Speechmatics
Dedicated medical models, with healthcare as a core segment. Speechmatics Enhanced Medical records 3.6% MER on the use-case table, second to Medical Mode. In the Medical Mode research benchmark, Enhanced medical averaged 9.7% WER, 6.1% medical-term miss and 11.1% drug-name miss in batch. In streaming it recorded 9.6% WER and 10.9% drug-name miss, lower than AssemblyAI with Medical Mode, at 4.71 s p90 time-to-final. Ask for its BAA terms in writing before you pilot.
3. Deepgram: Nova-3 Medical
A separate medical model that you route clinical audio to, for pre-recorded and streaming. It records 4.7% MER on the use-case table. In the Medical Mode research benchmark, nova-3-medical averaged 11.2% WER, 9.7% medical-term miss and 26.6% drug-name miss in batch, and 29.4% drug-name miss in streaming at 0.87 s p90 time-to-final. Terminology handling is keyterm prompting, which means lists to maintain per specialty, per formulary and per customer. BAAs are for Enterprise customers. The full head-to-head is in AssemblyAI vs Deepgram for medical transcription.
4. Amazon Transcribe Medical
A dedicated medical endpoint inside AWS, with HealthScribe layered on top for structured note output. It records 8.7% MER on the use-case table, the weakest of the dedicated medical options measured there. The BAA is accepted through AWS Artifact. For a workflow where a human reviews everything anyway, that accuracy may be acceptable; for an ambient scribe feeding an LLM, it's a problem.
5. Google Cloud Speech-to-Text and MedASR
Google Cloud offers medical conversation and medical dictation models, covered under the Google Cloud BAA. Google Medical Conversation records 24.4% MER on the use-case table. In the Medical Mode research benchmark, Google Gemini 3.5 Transcribe averaged 10.3% WER, 6.4% medical-term miss and 13.4% drug-name miss in batch. Separately, Google released MedASR, an open-weight, English-only medical speech model aimed at dictation, which you host and run yourself. It isn't in the use-case table, so measure it on your own audio.
6. ElevenLabs Scribe v2
A general transcription model with a medical variant, Scribe v2 Medical. It isn't in the use-case MER table. In the Medical Mode research benchmark, Scribe v2 Medical averaged 11.3% WER, 5.3% medical-term miss and 11.1% drug-name miss in batch. Scribe v2 Realtime finalized at 0.80 s p90 in streaming, with 13.5% WER and 19.6% drug-name miss. On pre-recorded diarization error rate, ElevenLabs Scribe v3 records 12.0% against 10.7% for AssemblyAI-STT-3.7. ElevenLabs signs BAAs through its sales team.
7. Microsoft Azure AI Speech
Healthcare capability comes through custom speech models and custom vocabulary rather than a single medical endpoint, under Microsoft's BAA for Azure services. It isn't in the use-case MER table. Expect more configuration work than a one-parameter medical option.
8. NVIDIA Riva
Not a hosted API: a toolkit you deploy and train. You own the model quality, and accuracy is whatever your training data supports. It fits air-gapped environments and teams with ML engineering capacity who want no external dependency. It doesn't fit a team that needs to ship a scribe this quarter.
9. Microsoft Dragon Copilot and Abridge
These are products clinicians use, not APIs developers build on. Microsoft Dragon Copilot combines Dragon Medical One dictation with DAX ambient listening; Abridge is an enterprise ambient scribe. If you're a health system buying documentation software, they're your shortlist. If you're building a clinical voice product, they're your competitive set.
Veterinary speech-to-text
Veterinary audio breaks human-medicine models in a specific way: they swap veterinary drugs for similar-sounding human ones, so a model can hear Zydax and write Xanax. Vet Mode, in beta, applies the same post-processing approach to veterinary vocabulary such as Cerenia, carprofen, Heartgard, trilostane and Antisedan. You enable it with domain: "veterinary-v1" on pre-recorded audio (AssemblyAI-STT-3.7 and Universal-2) and on streaming (AssemblyAI-STT-3.7-Realtime, AssemblyAI-STT-3.7-Realtime-Agent and Universal-3.6 Pro Realtime); the pre-recorded and streaming docs cover setup. See ambient scribes for veterinary and legal for the workflow.
A note on medical dictation
If your use case is front-end dictation, two AssemblyAI facts matter. The Dictation API runs speech-to-text and an LLM cleanup pass in one call and returns both the verbatim and the cleaned text. Medical Mode isn't available on it, so the 3.2% MER figure applies to the pre-recorded and streaming APIs, not to dictation. The Dictation API does take keyterms_prompt for drug names.
Getting started
Pre-recorded transcription with Medical Mode, speaker labels and a short keyterm list. Pre-recorded takes speech_models as a list:
import os
from assemblyai.prerecorded.v2 import Transcriber, TranscriptionConfig
config = TranscriptionConfig(
speech_models=["assemblyai-stt-3-7", "universal-2"],
domain="medical-v1",
speaker_labels=True,
keyterms_prompt=["metoprolol", "lisinopril", "atorvastatin"],
)
transcriber = Transcriber(api_key=os.environ["ASSEMBLYAI_API_KEY"])
transcript = transcriber.transcribe("encounter.wav", config)
if transcript.status == "error":
raise RuntimeError(transcript.error)
for utterance in transcript.utterances:
print(f"Speaker {utterance.speaker}: {utterance.text}")Streaming takes speech_model (singular) and connects to wss://streaming.assemblyai.com/v3/ws. This example replays a 16 kHz mono WAV in 100 ms chunks on AssemblyAI-STT-3.7-Realtime, with speaker labels and longer turn-silence settings for clinical pauses:
import os
import time
import wave
from assemblyai.streaming.v3 import (
RealTimeError,
RealTimeEvents,
RealTimeParameters,
RealTimeTranscriber,
RealTimeTranscriberOptions,
SpeakerRevisionEvent,
TurnEvent,
)
def on_turn(client: RealTimeTranscriber, event: TurnEvent):
if event.end_of_turn and event.transcript:
print(f"[{event.speaker_label}] {event.transcript}")
def on_revision(client: RealTimeTranscriber, event: SpeakerRevisionEvent):
# End-of-session pass: corrected speaker labels for earlier turns.
for revision in event.revisions:
print(f"turn {revision.turn_order} -> {revision.speaker_label}")
def on_error(client: RealTimeTranscriber, error: RealTimeError):
print(f"Error: {error}")
client = RealTimeTranscriber(
RealTimeTranscriberOptions(terminate_timeout=30.0),
api_key=os.environ["ASSEMBLYAI_API_KEY"],
)
client.on(RealTimeEvents.Turn, on_turn)
client.on(RealTimeEvents.SpeakerRevision, on_revision)
client.on(RealTimeEvents.Error, on_error)
client.connect(
RealTimeParameters(
sample_rate=16000,
speech_model="assemblyai-stt-3-7-realtime",
domain="medical-v1",
speaker_labels=True,
min_turn_silence=800,
max_turn_silence=2000,
)
)
# Replay a 16 kHz, mono, 16-bit WAV in 100 ms chunks, paced like live audio.
with wave.open("encounter.wav", "rb") as wav:
frames_per_chunk = wav.getframerate() // 10
while chunk := wav.readframes(frames_per_chunk):
client.stream(chunk)
time.sleep(0.1)
# Terminate flushes the last turn and waits for the Termination event.
client.disconnect(terminate=True)Register handlers with the RealTimeEvents enum; passing the TurnEvent class to client.on() registers nothing. Streaming runs on AssemblyAI-STT-3.7-Realtime.
Toggle Medical Mode and diarization in the playground and read the difference in the drug names yourself.
PHI, BAA, and deployment
Every option on this list can get through a security review with the right paperwork. The differences are in the specifics.
AssemblyAI is considered a business associate under HIPAA and offers a standard Business Associate Addendum (BAA) for customers processing PHI. Paid accounts can sign it self-serve from the dashboard; see the BAA FAQ. Supporting controls include PHI redaction across audio and transcripts and SOC 2 Type 2. EU data residency is available at api.eu.assemblyai.com.
The three questions to ask every vendor, in writing: Does redaction cover the audio or only the transcript? Is my audio retained, and for how long? Is it used for model training? Vendors answer these very differently, and the answers rarely appear on the pricing page.
Where teams get this wrong
For the organizational side (rollout, clinician adoption and the economics), see the overview of AI medical transcription in healthcare. On the technical side, these are the patterns that recur:
- Benchmarking on the wrong metric. A WER bakeoff will pick a model that's worse at drug names. Score entities.
- Benchmarking on the wrong audio. Clean read speech tells you nothing about far-field exam-room capture. Use your own recordings, at your own mic distances, with your own accents.
- Ignoring diarization until integration. Speaker attribution failures surface late and are expensive to work around. Test early, on short turns.
- Treating compliance as a final checkbox. Start the BAA conversation in parallel with the technical eval. It's rarely the blocker, but it's often the delay.
- Overloading keyterm lists. On long recordings, keep keyterms to about 200 or fewer; very long lists add false insertions. Free-text context about the encounter usually does more.
What the next comparison will look like
Ranked lists like this one have a short shelf life, and not because models get faster. The axis of comparison is moving.
Entity accuracy on clean benchmarks is converging at the top of the field; the top two on the use-case table are 0.4 points apart. What isn't converging is how well each system uses what you already know: the patient's history, the clinician's specialty, the prior encounter. In internal healthcare testing, passing a prior-visit note cut missed medical terms by 31%. The gap between a team that feeds context well and one that doesn't will soon be larger than the gap between any two vendors on this page.
Tell us your use case, audio profile and compliance constraints, and get help designing an eval that produces a defensible answer.
Frequently Asked Questions (FAQs)
What is the best medical speech-to-text?
It depends on whether you're buying software or building it. For developers building a clinical product, Universal-3.5 Pro with Medical Mode records a 3.2% Missed Entity Rate on AssemblyAI's published medical benchmark, ahead of Speechmatics, Deepgram, AWS and Google. For a health system buying finished documentation software, the comparison set is products like Microsoft Dragon Copilot and Abridge.
Medical dictation software vs speech-to-text API: which do I need?
Dictation software is a finished app a clinician talks into; a speech-to-text API is the layer developers use to build dictation apps, ambient scribes and clinical voice agents. Buy dictation software if you're a provider organization documenting your own visits. Use an API if documentation is part of a product you sell or your workflow doesn't fit a standard template.
Which speech-to-text is most accurate for medical terminology?
On Missed Entity Rate, the metric that scores clinically meaningful words rather than all words equally, Universal-3.5 Pro with Medical Mode records 3.2%, followed by Speechmatics Enhanced Medical at 3.6%, Deepgram Nova-3 Medical at 4.7%, AWS Transcribe Medical at 8.7% and Google Medical Conversation at 24.4%. Those figures come from one published table. Run your own audio before deciding. For drug names specifically, a separate batch benchmark of six systems across 9 sets in 4 languages found Medical Mode had the lowest average drug-name miss rate, 10.7%, against 11.1% for Speechmatics and ElevenLabs Scribe v2 Medical.
Is there speech-to-text for veterinary notes?
Yes. AssemblyAI's Vet Mode is in beta and is enabled with domain: "veterinary-v1" on pre-recorded and streaming audio. It corrects veterinary drug and procedure names that general models misspell or swap for similar-sounding human drugs, such as writing Xanax for Zydax. It runs on AssemblyAI-STT-3.7 and Universal-2 for pre-recorded audio, and on AssemblyAI-STT-3.7-Realtime, AssemblyAI-STT-3.7-Realtime-Agent and Universal-3.6 Pro Realtime for streaming.
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 accounts can review and sign it self-serve from the dashboard, without a sales call.
Does medical speech-to-text work with EHRs?
Speech-to-text APIs return transcripts; writing to the EHR is your integration layer's job, or your scribe vendor's. What the transcription layer needs to provide is accurate medical entities and reliable speaker attribution, so downstream extraction can map patient statements and clinician instructions into the right fields. Finished products such as Microsoft Dragon Copilot and Abridge ship their own EHR integrations.