How to use Voice AI for healthcare market research

Picture the end of a study. You've recorded sixty in-depth interviews with nephrologists and oncologists across three countries, plus a dozen payer advisory boards. Every one of those conversations is dense with drug names, dosing regimens, off-label indications, competitor brand names and the kind of hedged, half-finished sentences clinicians use when they're thinking out loud. Somewhere in there is the answer your client paid for. Right now it's locked inside audio files.

The traditional unlock is a human transcriptionist and a coding team, and it costs weeks. The modern unlock is a transcription API — but only if the model can actually hear the vocabulary. That's the part most teams get wrong. They pick a general-purpose speech model, watch it render "empagliflozin" as "empire glyphosate," and conclude that voice AI isn't ready for healthcare work. The model wasn't ready. The category is.

With Universal-3.5 Pro and Medical Mode switched on, AssemblyAI hits a 3.2% Missed Entity Rate on medical entities — the lowest across benchmarked providers. It runs at $0.36/hr combined. Sixty hours of interviews cost about twenty-two dollars to transcribe. This post walks through how to build the pipeline that turns those hours into coded, searchable, quotable insight.

What healthcare market research actually needs from voice AI

Market research audio is its own beast. It isn't a clean dictation, and it isn't a two-person clinical encounter. It's a moderator plus one to eight respondents, often on a video call, sometimes with a translator on the line, always with someone eating lunch off-mic.

Four requirements fall out of that.

Entity-level accuracy, not word-level accuracy

A 95% word accuracy score sounds excellent until you notice the 5% it missed was every product name in the study. Word Error Rate treats "the" and "tirzepatide" as equally weighted tokens. They are not equally weighted to you. The metric that matters for healthcare research is Missed Entity Rate — how often the model drops or mangles a drug, condition, procedure or dosage. That's what Medical Mode is built to move.

Reliable speaker separation

An uncoded transcript where the moderator's prompt and the respondent's answer are fused into one paragraph is worse than useless — it manufactures quotes that nobody said. You need diarization you can trust at the turn level.

Multilingual handling that survives code-switching

European and LATAM studies rarely stay in one language. A German KOL will drop English trial acronyms mid-sentence. A Mexican respondent will say the brand name in English and the indication in Spanish. A model that requires you to declare a language up front will mangle every one of those switches.

PHI discipline you can hand to a compliance reviewer

Even when a study is de-identified by design, respondents mention patients. That's PHI the moment it lands in your transcript. You need redaction, and you need a signed agreement covering the processor.

Why general-purpose transcription breaks on clinical language

General models are trained to be plausible. When they hear an unfamiliar polysyllabic word, they substitute the nearest common-English phrase. In casual media that's a harmless artifact. In a healthcare study it silently rewrites your data — and it rewrites it in a way that's hard to spot on review, because the output reads fluently.

Medical Mode changes the decoding behavior rather than swapping the model. You keep Universal-3.5 Pro and add one parameter. The measured effect: 87% fewer entity errors and roughly 20% fewer missed medical entities compared with the base model running without it. The full methodology and the head-to-head numbers live on the benchmarks page.

Viable general transcription; expect manual QA on entity terms

Build the transcription pass

Start with the async API. Interview recordings are batch work — there's no reason to pay streaming rates for a file you recorded yesterday.

import requests, time

BASE = "https://api.assemblyai.com/v2"
HEADERS = {"authorization": "YOUR_API_KEY"}

payload = {
    "audio_url": "https://example.com/kol-interview-042.wav",
    "speech_models": ["universal-3-5-pro"],
    "domain": "medical-v1",
    "speaker_labels": True,
    "redact_pii": True,
    "redact_pii_policies": [
        "person_name",
        "date_of_birth",
        "phone_number",
        "email_address"
    ],
    "redact_pii_audio": True
}

job = requests.post(f"{BASE}/transcript", json=payload, headers=HEADERS).json()
print(job["id"])

Three things are doing the work there. speech_models: ["universal-3-5-pro"] picks the async flagship. domain: "medical-v1" turns on Medical Mode — that's the whole activation, no model switch. And redact_pii_audio bleeps the source audio as well as the text, which matters when your deliverable includes verbatim clips.

One caution on redaction policy for research work: redacting clinical conditions will strip the thing you're studying. Most research teams redact identity fields aggressively and leave clinical content intact, then handle clinical sensitivity at the reporting layer. Decide that deliberately rather than copying a default. The API docs list every policy.

Separating the moderator from the respondent

Diarization is the difference between a transcript and a dataset. Universal-3.5 Pro ships the most accurate diarization we've released, and the reason it works better on research audio is what it optimizes for: cpWER rather than DER. In plain terms, it's tuned to get the right words attributed to the right speaker, not just to draw tidy boundaries between speech regions.

That shows up in exactly the places research audio is hardest. Short turns — "Mm-hm," "Right," "Both, actually" — survive instead of being absorbed into the neighboring speaker. Overlapped speech, which is constant in group discussions, gets untangled instead of collapsed.

Turn speaker_labels on and you get utterance-level attribution you can pipe straight into a coding framework:

while (result := requests.get(f"{BASE}/transcript/{job['id']}", headers=HEADERS).json())["status"] != "completed": time.sleep(3)

for utt in result["utterances"]:
    print(f"[{utt['speaker']}] {utt['text']}")

For advisory boards and group discussions, streaming diarization supports revision across up to 10 speakers, which is the practical ceiling for a moderated panel anyway.

Multilingual studies without a language-per-file pipeline

Universal-3.5 Pro code-switches natively across 18 languages with no configuration. You don't declare a language, you don't route files to per-language endpoints, and you don't lose the English trial acronym a German respondent drops into a German sentence.

Medical Mode itself covers four languages — English, Spanish, German and French — across both pre-recorded and streaming. Those are two separate facts, and it's worth keeping them separate when you scope a study: the base model's code-switching is broad, and the medical entity boost applies to those four. For a study running in Japanese or Mandarin you still get strong general transcription; you just don't get the Medical Mode entity lift.

Contextual prompting is the step with the biggest payoff

Every study has its own private vocabulary — the molecule code that hasn't been branded yet, the internal name for a comparator arm, the three abbreviations your client uses that nobody else does. You already have documents containing all of it: the discussion guide, the screener, the client's briefing deck.

Feed that context to the model. In an internal healthcare test, supplying a patient's prior-visit note cut missed medical terms by 31%. The mechanism is the same for research: give the model the document that names the things it's about to hear, and it stops guessing.

This is the step teams skip because it feels like cheating, and it's the step with the biggest measured return. Build it into your study setup so it happens by default rather than when someone remembers.

From transcript to insight

A clean, attributed transcript is the input to analysis, not the analysis. Once you have it, you can run structured extraction over the corpus — theme identification, sentiment per respondent segment, mentions of each competitor product, unprompted versus prompted awareness.

The LLM Gateway lets you do that against the transcript without standing up a separate inference stack. Practically, the pattern that works is: extract per-interview into a rigid schema first, then aggregate across interviews. Asking a model to summarize sixty interviews at once produces something that reads well and cites nothing.

prompt = """You are coding a physician interview for a market research study.
Return JSON with keys: unprompted_brand_mentions, prescribing_barriers,
efficacy_perceptions, unmet_needs. Quote the respondent verbatim for each
item and include the utterance index. Do not infer beyond the transcript."""

# Send `prompt` plus the diarized transcript to the LLM Gateway,
# then aggregate the per-interview JSON across the study in your own code.

Two rules keep this defensible. Require verbatim quotes with utterance indices so every claim is traceable back to audio. And never let a model summarize across respondents in a single pass — aggregate deterministically yourself.

Live sessions and real-time analysis

Some research is better instrumented live — a moderator getting nudges during an interview, a back room watching a real-time transcript, a screener that routes respondents based on what they just said.

For that, Universal-3.5 Pro Realtime streams over wss://streaming.assemblyai.com/v3/ws at $0.45/hr base, and Medical Mode adds the same $0.15/hr for $0.60/hr combined.

import websockets; from urllib.parse import urlencode

CONFIG = {
    "speech_model": "universal-3-5-pro",
    "domain": "medical-v1",
    "mode": "max_accuracy",
    "voice_focus": "far-field",
    "speaker_labels": True,
    "prompt": "Interview with a nephrologist about SGLT2 inhibitors."
}

async def stream(mic):
    url = "wss://streaming.assemblyai.com/v3/ws?" + urlencode({k: str(v).lower()
if isinstance(v, bool) else v for k, v in CONFIG.items()})
    async with websockets.connect(
        url, additional_headers={"Authorization": "YOUR_API_KEY"}
    ) as ws:
        # Config travels in the query string above; v3 has no configure message
        async for chunk in mic:
            await ws.send(chunk)

Note that speech_model is singular for streaming while speech_models is plural for async — a small inconsistency that trips up everyone exactly once.

Three streaming options earn their keep in research settings. voice_focus takes near-field or far-field; pick far-field for a conference-room mic or a laptop across a table. mode takes min_latency, balanced or max_accuracy — for research, accuracy wins, since nobody is waiting on a sub-second reply. And prompt carries study context into the decode; across 20,000 voice agent files context cut WER by 10.2%, with detailed context cutting medical-term entity errors 43%. Turn detection defaults to min_turn_silence 128ms and max_turn_silence 1280ms on the balanced preset, which is what keeps a live transcript feeling synchronous.

PHI, consent and where the data lives

Compliance isn't the headline of a research program, but it's the thing that stops one. The relevant facts, stated plainly:

• AssemblyAI signs a Business Associate Addendum (BAA) for customers processing PHI, acting as a business associate under HIPAA. Details on the BAA FAQ and the BAA page.
• PHI redaction runs across both audio and transcripts, so verbatim clips in a deliverable can be redacted at the source.
• SOC 2 Type 2 covers the platform controls your client's security review will ask about.
• EU data residency is available at api.eu.assemblyai.com, and self-hosted deployment exists for programs that can't send audio out at all.

Get the BAA executed before the first file moves, not during fieldwork. It's a paperwork step, and paperwork steps expand to fill whatever slack you leave them.

What it costs

Pricing is simple enough to model on a napkin. Async transcription runs $0.21/hr on Universal-3.5 Pro, $0.36/hr with Medical Mode. Live streaming runs $0.45/hr base, $0.60/hr with Medical Mode. If you're building a conversational screener that talks back, the Voice Agent API is a flat $4.50/hr for one WebSocket replacing the whole STT plus LLM plus TTS chain. Current numbers are always on the pricing page.

The interesting shift isn't cost per hour, though — it's that you stop rationing. When transcription is effectively free, you transcribe everything: the pilot interviews, the screener calls, the sessions you'd have written off as unusable.

Where this goes

The pattern worth watching isn't better transcripts. It's studies that no longer have a distinct analysis phase at all. When live transcription is accurate enough on clinical vocabulary and cheap enough to run on every session, the coding frame starts updating during fieldwork instead of after it — which means the discussion guide for interview forty can be shaped by what came out of interview thirty-nine. That's a different research method, not a faster version of the current one. Teams like Heidi Health, Sully AI and Chapter are already running this class of pipeline on clinical and member audio, and the research side of healthcare is next. Building on accurate speech infrastructure now is what makes that possible later.

Frequently asked questions

What is the best speech-to-text API for medical transcription and healthcare research?

For work where drug names, conditions and procedures have to survive intact, the deciding metric is Missed Entity Rate. Universal-3.5 Pro with Medical Mode measures 3.2% MER — the lowest across benchmarked providers. Full methodology is on the benchmarks page. For research specifically, diarization quality matters just as much, since misattributed quotes are worse than missing ones.

How does AssemblyAI handle HIPAA and PHI?

AssemblyAI signs a Business Associate Addendum (BAA) for customers processing PHI, and operates as a business associate under HIPAA. Alongside that: PHI redaction across audio and transcripts, SOC 2 Type 2, EU data residency, and a self-hosted deployment option. Start at the BAA FAQ.

How does AssemblyAI capture medical jargon that general models miss?

Two mechanisms stack. Medical Mode (domain: "medical-v1") biases decoding toward clinical vocabulary and delivers 87% fewer entity errors than the base model without it. On top of that, contextual prompting lets you supply study-specific documents — a discussion guide, a briefing deck, a prior-visit note — which cut missed medical terms by 31% in an internal healthcare test.

Which languages does Medical Mode support for international studies?

Medical Mode covers English, Spanish, German and French, for both pre-recorded and streaming audio. Separately, the base Universal-3.5 Pro model code-switches natively across 18 languages with no configuration, so mixed-language interviews transcribe correctly even outside those four.

Does AssemblyAI automatically redact patient PII from transcripts?

Not by default — PII and PHI redaction is an opt-in request parameter, and it applies to the audio as well as the text. That second part matters for research deliverables that include verbatim clips. You choose which policies apply, which is a decision worth making deliberately: redacting clinical conditions will strip the substance of a healthcare study.

Can I analyze live interviews as they happen?

Yes. Universal-3.5 Pro Realtime streams over wss://streaming.assemblyai.com/v3/ws with turn detection defaulting to min_turn_silence 128ms and max_turn_silence 1280ms on the balanced preset, diarization with revision for up to 10 speakers, and voice_focus tuned for near-field or far-field microphones. See the Universal-3.5 Pro Realtime post and the healthcare solutions page for build patterns.

‍