Medical transcription services: how to choose one, and when to build instead

An orthopedic group with 40 providers runs about 320 encounters a day. At six minutes of dictation each, that's 32 hours of audio daily. Whoever handles that audio — an outsourced transcription vendor, an AI service, or software the group builds itself — is making a decision that touches every note in the chart, every coded claim, and every minute of physician time after clinic closes.

The decision usually gets framed as a vendor selection. It's really an architecture decision. Traditional transcription services price by volume and hand you finished documents. AI services price by volume too but return results in minutes. And a speech-to-text API prices by the hour of audio — $0.36/hr for Universal-3.5 Pro with Medical Mode enabled — and hands you a transcript you have to do something with.

Those three options have wildly different cost curves, different accuracy characteristics, and different implications for what your organization can build later. This post lays out how they differ, what to actually evaluate, and how to read the accuracy claims that every vendor in this market puts on their homepage.

What a medical transcription service does

At minimum, a medical transcription service converts clinical audio into text: physician dictations, patient encounters, procedure notes, consult letters, telehealth visits, care coordination calls. Most also promise some combination of formatting into a note template, turnaround time guarantees, quality assurance review, and delivery into an EHR.

The category has widened considerably. "Medical transcription service" in 2020 meant a company with transcriptionists. Today it covers pure-software offerings, ambient scribing products, and raw APIs — with very different assumptions about who does the review and where the PHI lives.

The three service models

Human transcription services

Certified medical transcriptionists or editors listen to the audio and type the note. Pricing is typically per audio minute or per 65-character line. Turnaround runs from a few hours to a day or more, and cost scales linearly with volume — every additional encounter costs the same as the last one.

The genuine strength is judgment. A human transcriptionist notices when a dictation contradicts itself, flags an ambiguous dose, and handles a heavy accent or a badly-recorded phone dictation that would defeat most software. The genuine weakness is that you cannot grow volume without growing spend, and turnaround is bounded by human availability.

AI transcription services

Software transcribes the audio, usually in minutes, with no human in the loop unless you add one. Cost per encounter drops dramatically and turnaround becomes near-instant. Quality depends almost entirely on how well the underlying model handles clinical vocabulary, which is the thing vendors are least specific about.

This is where the market is moving, and where accuracy diligence matters most. A general-purpose speech model wrapped in a healthcare-branded UI will transcribe conversational English beautifully and lose drug names. That's not a hypothetical — it's the default failure mode.

Hybrid: AI draft plus human review

Software produces the draft; a human editor reviews before release. You get most of the cost reduction and keep a quality gate. In practice, hybrid is where regulated, high-stakes documentation lands — surgical notes, pathology, anything feeding a coded claim — and it's a sensible place to start even if you plan to remove the human step later.

The economics of hybrid depend entirely on how good the draft is. Reviewing a transcript with a 3% entity miss rate is a fast skim. Reviewing one with a 24% miss rate is retyping. That single variable determines whether hybrid saves money or just adds a step.

Building on an API

The fourth option most buyers don't consider: use a speech-to-text API directly and build the workflow you actually want. You pay per hour of audio, you own the transcript and the interface, and you can integrate with your EHR on your own terms. It's the right answer for health tech companies whose product is the workflow, and increasingly for large systems with engineering capacity.

How to read accuracy claims

Every vendor in this category advertises an accuracy percentage. Almost none of them tell you what it measures, and the difference matters enormously.

Word error rate flatters everyone

A "98% accurate" claim usually means 2% word error rate: two of every hundred words are wrong. But function words dominate any transcript. Getting "the," "and," and "patient" right while missing two medication names produces an excellent WER and a dangerous note. Word error rate is a fine metric for tracking a model's general quality and a poor one for deciding whether a clinical transcript is usable.

Missed Entity Rate is the number to ask for

Missed Entity Rate scores only clinically meaningful tokens — drugs, dosages, conditions, procedures, anatomy. Universal-3.5 Pro with Medical Mode posts a 3.2% MER on our benchmark, the lowest across the providers evaluated, against roughly 8.7% for Deepgram Nova-3 Medical and roughly 24.4% for AWS Transcribe Medical. Methodology is published on our benchmarks page.

Turning Medical Mode on delivers about 20% fewer missed medical entities than the base model, and 87% fewer entity errors overall. If a vendor can't give you an entity-level figure, ask them to run your audio and count the drug names themselves. For a deeper treatment of how these numbers are constructed, see our post on AI medical transcription accuracy; for why specialist vocabulary breaks models in the first place, see how accurate AI transcripts are on technical and medical terms.

What else to evaluate

Speaker separation

For dictation, one speaker is all you need. For patient encounters, attribution is the whole game — "I stopped taking the lisinopril" is a different clinical fact depending on who said it. Universal-3.5 Pro's diarization is optimized for cpWER rather than diarization error rate, which means it's scored on whether the words land under the right speaker, not just on how much audio was roughly assigned correctly. It's the most accurate diarization we've shipped and it holds up on short turns and overlapped speech — the two things that actually happen in exam rooms.

Language coverage

Two different questions hide inside "do you support Spanish?" Universal-3.5 Pro code-switches natively across 18 languages with no configuration, so a patient who switches between English and Spanish mid-sentence transcribes correctly without a language hint. Medical Mode's clinical domain adaptation covers English, Spanish, German, and French. Ask any vendor both questions separately — general language support and clinical-vocabulary support are not the same coverage list.

PHI handling and contracting

This is the shortest conversation to have early, because it can eliminate vendors. AssemblyAI signs a Business Associate Addendum (BAA) for customers processing PHI and operates as a business associate under HIPAA. PHI redaction is available across both audio and transcripts, the platform is SOC 2 Type 2 audited, and self-hosted deployment and EU data residency are available where residency rules apply. The specifics are in the BAA FAQ and the BAA terms.

Ask every candidate: will you sign a BAA, where does audio live and for how long, can I delete it on demand, and is there a redaction step before storage.

Turnaround and how it's measured

"Same-day" from a human service and "under a minute" from an API are not comparable promises, and neither is what the physician experiences. What matters is whether the note is ready before the clinician has moved on. Async transcription finishing in minutes clears that bar for almost every documentation workflow. Streaming only matters if something in your product needs the words while the person is still talking.

Integration surface

A service that emails you documents is a workflow dead end. Ask what you get programmatically: transcript with word-level timings, speaker labels, confidence scores, structured entities. Those are the raw materials for coding support, quality review, and analytics later. If a vendor won't expose the transcript, you're buying a document, not a data asset.

Pricing model, not just price

Per-seat pricing rewards you for low utilization and punishes growth. Per-minute pricing scales with volume. Per-hour-of-audio API pricing is the most predictable of the three because it maps directly to something you can forecast. Our full breakdown is on the pricing page.

Building it yourself: what's actually involved

If you have engineers, the API path is less work than most teams assume. Here's a complete clinical transcription call:

import assemblyai as aai

aai.settings.api_key = "YOUR_API_KEY"

config = aai.TranscriptionConfig(
    speech_models=["universal-3-5-pro"],
    domain="medical-v1",
    speaker_labels=True,
    redact_pii=True,
    redact_pii_audio=True, redact_pii_policies=[aai.PIIRedactionPolicy.person_name, aai.PIIRedactionPolicy.date_of_birth, aai.PIIRedactionPolicy.phone_number],
)

transcript = aai.Transcriber(config=config).transcribe(
    "https://storage.example.org/encounters/2026-08-27-0913.wav"
)

for utterance in transcript.utterances:
    print(f"Speaker {utterance.speaker}: {utterance.text}")

That single request gives you a diarized clinical transcript with PHI redacted in both the text and the audio. Note the plural speech_models field — async transcription takes a list; streaming uses the singular speech_model.

The accuracy lever most teams miss is contextual prompting. Passing a patient's prior-visit note along with the audio cut missed medical terms by 31% in an internal healthcare test, because the model gets that patient's actual medication list and diagnoses before it starts decoding. If your system already has the chart open when recording starts, that context costs you nothing to supply and it beats maintaining a custom vocabulary file that goes stale.

For live workflows, streaming uses the v3 WebSocket endpoint at wss://streaming.assemblyai.com/v3/ws with Universal-3.5 Pro Realtime at $0.45/hr base, $0.60/hr with Medical Mode. Full request and response reference is in the docs.

Which model fits which organization

A few patterns hold up reliably.

Small practices with modest volume and no engineering staff should buy a finished product. The per-encounter premium is real but the alternative is a project nobody has time to run.

Large systems with high volume should look hard at the API path for the bulk workload and keep a hybrid review step for the documentation types that feed claims. The cost difference at 30-plus hours of audio a day is large enough to fund the engineering.

Health tech companies building for clinicians should build on an API, full stop. If documentation is your product, the workflow is your differentiation, and you can't differentiate on a vendor's interface. Teams working this way on AssemblyAI's speech-to-text include Sully AI, Heidi Health, Deepscribe, Knowtex, Commure, and Magentus Healthcare, along with organizations like Chapter and NMDP applying the same stack outside the exam room.

Two adjacent workloads are worth naming, because buyers often discover mid-evaluation that they actually need one of them. If you want documentation generated from the natural doctor-patient conversation rather than from dictation, that's an ambient AI scribe, and the acoustic requirements are different — far-field capture, robust diarization, overlap handling. If you want the system to talk back to patients, that's a telehealth triage voice agent, which is a real-time problem with latency requirements a transcription service will never meet.

Where this goes

The transcription service category is being quietly dissolved. What health systems bought for thirty years was a document-production service; what they're buying now is a data pipeline that happens to produce documents. The interesting consequence is that the transcript stops being the deliverable. Once entity accuracy is high enough to trust, the same audio feeds coding support, quality measurement, and prior-authorization evidence without a second capture step. Organizations picking a vendor purely on cost per minute are optimizing the one variable that will matter least in three years. Pick on whether you get the transcript, the entities, and the ability to build on top of them.

Frequently asked questions

How much do medical transcription services cost?

It depends on the model. Human and hybrid services price per audio minute or per line, so cost scales directly with volume. AI services usually price per minute or per clinician seat. If you build on an API, you pay per hour of audio: $0.21/hr for Universal-3.5 Pro, plus $0.15/hr for Medical Mode, so $0.36/hr combined for async clinical transcription. Streaming is $0.45/hr base and $0.60/hr with Medical Mode. See pricing for details.

What accuracy should I expect from a medical transcription service?

Ask for an entity-level number rather than a general accuracy percentage. Universal-3.5 Pro with Medical Mode posts a 3.2% Missed Entity Rate — the lowest among the providers on our benchmark, where Deepgram Nova-3 Medical lands near 8.7% and AWS Transcribe Medical near 24.4%. A vendor quoting only word error rate is quoting the metric that hides medication errors.

What is AssemblyAI's Medical Mode?

It's a clinical domain adaptation for our flagship models, enabled with one parameter: domain: "medical-v1". No model switch, no separate endpoint, no vocabulary list to maintain. It adds $0.15/hr and cuts entity errors 87% versus the same model without it. It works on both async transcription and streaming.

How does AssemblyAI handle HIPAA and PHI?

AssemblyAI signs a Business Associate Addendum (BAA) for customers processing PHI and acts as a business associate under HIPAA. Alongside that: PHI redaction across audio and transcripts, SOC 2 Type 2 audited infrastructure, and self-hosted or EU-resident deployment options. Start with the BAA FAQ.

Can a medical transcription service handle accents and multiple languages?

Universal-3.5 Pro handles accented English well and code-switches natively across 18 languages without configuration, which covers the common case of a patient moving between two languages in one sentence. Medical Mode's clinical vocabulary adaptation covers English, Spanish, German, and French — a shorter list than the base model's language range, so check both.

Do I need real-time transcription or is batch enough?

Batch is enough for almost all documentation. The audio uploads after the visit, the transcript comes back in minutes, and it's more accurate because the model sees the full recording. Choose streaming only when something in the experience needs words while the person is still speaking — a live-forming note, or a voice agent holding a conversation.

‍