Introduction

A clinical transcript can be 99% right and still be dangerous. If "15 mg" becomes "50 mg", or a dog's carprofen turns into ibuprofen, someone has to catch it before it reaches the record. AssemblyAI went after that last 1% today.

On November 10 (UTC), AssemblyAI released an updated Medical Mode for both recorded and streaming audio, and a new Vet Mode in beta for veterinary terms. Both are closed, API-only features that sit on top of AssemblyAI's Universal speech-to-text models. You turn them on with a single domain parameter, and you can try them now in the AssemblyAI playground or API.

What Shipped

  • Features: updated Medical Mode (domain: "medical-v1") and Vet Mode beta (domain: "veterinary-v1").
  • Vendor: AssemblyAI, announced in a blog post and on
    .
  • Modality: speech-to-text with a correction pass tuned for clinical language. It finds drug names, procedures and conditions the model heard but spelled wrong, and fixes them using the words around them.
  • Recorded audio: Universal-3.5 Pro and Universal-2.
  • Streaming: Universal-3.6 Pro Realtime, Universal-3.5 Pro Realtime, and Universal-Streaming English and Multilingual.
  • Medical Mode languages: English, Spanish, German and French.
  • Price: Medical Mode is a $0.15 per hour add-on. The blog doesn't list a separate price for the Vet Mode beta.
  • Already using it? Anyone who already sends domain: "medical-v1" gets the update automatically. No code changes needed.
  • Access: closed. Use it through the API or the playground. AssemblyAI signs a HIPAA BAA, is SOC 2 Type 2 certified, and offers PHI redaction, zero data retention and on-prem deployment.

The Numbers

AssemblyAI tested on nine medical datasets in the four supported languages. It ran five other vendors in their medical modes where they had one. No system got custom vocabulary or keyterms. These are AssemblyAI's own results, so treat them as a starting point.

Recorded audio (average over nine sets)

  • AssemblyAI + Medical Mode: 9.5% WER, 5.2% medical-term miss rate, 10.7% drug-name miss rate.
  • Speechmatics (medical domain): 9.7% WER, 6.1% medical-term misses, 11.1% drug-name misses.
  • ElevenLabs Scribe v2 Medical: 11.3% WER, 5.3% medical-term misses, 11.1% drug-name misses.
  • Soniox stt-async-v6: 10.2% WER, 6.1% medical-term misses, 16.7% drug-name misses.
  • Gemini 3.5 Transcribe: 10.3% WER, 6.4% medical-term misses, 13.4% drug-name misses.
  • Deepgram nova-3-medical: 11.2% WER, 9.7% medical-term misses, 26.6% drug-name misses.

Compared with a base AssemblyAI model, Medical Mode misses up to 13% fewer medical terms, with no rise in wrongly inserted terms. On the older Universal-2 model, the cut is 27%.

Streaming (time-to-final is the 90th percentile)

  • Universal-3.6 Pro Realtime + Medical Mode: 10.4% WER, 6.3% medical-term misses, 13.0% drug-name misses, 1.28 s.
  • Same model without Medical Mode: 10.4% WER, 6.7% medical-term misses, 15.7% drug-name misses, 0.63 s, the fastest in the test.
  • Speechmatics (medical): 9.6% WER and 10.9% drug-name misses, but 4.71 s to finalize.
  • Soniox stt-rt-v6: 10.2% WER, 16.4% drug-name misses, 5.18 s.
  • ElevenLabs Scribe v2 Realtime: 13.5% WER, 19.6% drug-name misses, 0.80 s.
  • Deepgram nova-3-medical: 13.2% WER, 29.4% drug-name misses, 0.87 s.
  • Gemini 3.5 Transcribe Live: 16.2% WER, 25.2% drug-name misses, 1.32 s.

So Speechmatics and Soniox still edge it on raw streaming accuracy, but they commit words about five seconds behind the speaker. Among systems that finalize within about a second, Medical Mode leads on all three error measures. Turning it on costs you roughly 0.65 seconds in exchange for fewer missed drug names.

One timing note: AssemblyAI compared against Scribe v2. ElevenLabs launched Scribe v3 and Scribe v3 Live a few hours later, along with a medical version of Scribe v3 Live inside ElevenAgents, so this table is already a little out of date.

Vet Mode, in Beta

Many vet drugs sound like human drugs, and a general model tends to write the word it has seen more often. Vet Mode runs the same kind of correction pass, tuned for veterinary language. It fixes misspellings and wrong-but-real words. Examples from early customers:

  • "Syringea" → Cerenia
  • "Xanax" → Zydax
  • "ibuprofen" → carprofen
  • "Tri-Lysteine" → Trilostane
  • "Harcourt" → Heartgard
  • "antecedent" → Antisedan

It runs on the same recorded and streaming models as Medical Mode. See the recorded and streaming docs.

How to Try It

For recorded audio, add the domain to a normal transcript request:

import requests

response = requests.post(
    "https://api.assemblyai.com/v2/transcript",
    headers={"authorization": "YOUR_API_KEY"},
    json={
        "audio_url": "https://example.com/clinical-visit.mp3",
        "speech_models": ["universal-3-5-pro"],
        "domain": "medical-v1",
    },
)
print(response.json()["id"])

For streaming, pass it as a connection parameter:

wss://streaming.assemblyai.com/v3/ws?sample_rate=16000&speech_model=universal-3-6-pro&domain=medical-v1

Swap in veterinary-v1 for Vet Mode. The streaming Medical Mode docs cover the full setup.

Pro Tip

AssemblyAI says the lift depends on your audio and vocabulary. Run the same files with and without the domain flag, then score only the terms your users care about, like drug names, doses and conditions. For a live agent where every half second matters, you might leave it off for chit-chat and keep it on for intake and refill calls.

Where It Fits

This is for ambient scribes, clinical dictation, telehealth recordings, and voice agents that book appointments or take refill requests. It pairs well with speaker labels and PHI redaction. If you're comparing medical speech-to-text, also look at Speechmatics Oak. If you need open weights you can host yourself, NVIDIA's Nemotron 3.5 Transcribe is the one to test. And if you want the voice side of the same agent, AssemblyAI's TTS-1 shipped two weeks ago.

Conclusion

This isn't a new model. It's a one-line flag that targets the words that matter most in a clinical note. The streaming trade-off is honest and easy to test, and Vet Mode fixes a real problem that general models keep getting wrong. With ElevenLabs shipping Scribe v3 the same day, medical speech-to-text is moving fast, so I'd re-run your own benchmark before you pick a winner.

—Anabel ♡