Skip to content
CCompareHealthcareAPI

Buyer guide · Medical speech-to-text API

Medical speech-to-text API: what to evaluate, and what a transcript never gives you

A transcript is about 20% of a clinical documentation feature. This guide covers how to pick the engine and how to size the other 80%.

Written by Compare Healthcare API Editorial DeskReviewed by Technical ReviewLast reviewed Rubric v1.0

Short answer

What should you evaluate in a medical speech-to-text API?

Six things decide the choice: medical vocabulary and entity handling, speaker diarization quality on two-party conversations, streaming latency and partial-result behaviour, HIPAA posture (BAA, retention, whether audio trains models), published per-minute pricing, and self-serve access to a sandbox. Deepgram leads our medical speech-to-text ranking on developer experience and price transparency; AWS Transcribe Medical and Azure AI Speech are strong when you are already committed to that cloud. None of them return a clinical note — that layer remains yours.

Cite as: Medical speech-to-text API buyer guide, Compare Healthcare API, last reviewed 2026-09-01.

Key findings

  • Word error rate is the most cited and least useful number in this category, because no vendor publishes a reproducible corpus, audio-condition set or reference-transcript protocol.
  • Diarization quality on natural two-party clinical conversation is usually the real accuracy constraint, not vocabulary coverage.
  • Streaming and batch are different products with different failure modes; pick one deliberately rather than adopting whichever the demo used.
  • Cloud-native engines are cheap to start and expensive to leave: your documentation layer ends up shaped by their output format.
  • If you choose ASR, budget a permanent clinical evaluation function, not a one-off accuracy test.

Medical vocabulary is table stakes; structure is the hard part

Every serious engine now handles drug names, dosages, anatomy and common abbreviations at a level adequate for clinical dictation. Differences that matter show up elsewhere: how the engine behaves on overlapping speech, on a patient with an unfamiliar accent, on a noisy room, and on the long stretches of non-clinical conversation that ambient capture inevitably includes.

A transcript, however good, still contains none of the decisions a note requires: what is history versus assessment, what belongs in the chart at all, which utterances are patient-reported and which are clinician conclusions. Those decisions are the documentation layer, and buying ASR means building it.

Streaming versus batch

Streaming returns partial hypotheses during the encounter and is required if your UI shows live text or if a clinician acts on the transcript mid-visit. It brings connection management, partial-result reconciliation, reconnection handling and a materially harder testing story.

Batch processes a completed recording and is simpler, cheaper to operate and usually more accurate because the model sees full context. For most post-encounter documentation products, batch is the correct default and streaming is a feature to add when a customer actually needs it.

HIPAA posture for an ASR vendor

The ASR vendor becomes a business associate the moment encounter audio reaches it, so a signed BAA is a precondition for production traffic. Beyond the BAA, three settings decide whether you survive a health system's security review: how long audio and transcripts are retained by default, whether retention can be set to zero, and whether your data can be used to train the vendor's models.

Hyperscaler services publish HIPAA-eligible service lists and BAA terms, which is genuinely useful evidence; it also means your compliance story becomes an account-configuration story, and misconfiguration is your liability, not theirs.

What this path costs you

Our integration effort model puts an ASR-plus-own-documentation-layer path at 12 to 30 engineering weeks to a production-quality clinical note, versus 4 to 8 weeks for an embeddable scribe API. The larger cost is ongoing: prompt and template maintenance, clinical review of drafts, regression testing when a model changes underneath you, and ownership of clinical safety questions when a note omits something material.

Choose this path when documentation is your differentiator, when you have clinical staff to evaluate output, or when regulatory or contractual constraints require you to control the whole pipeline. Otherwise, buying the documentation layer is cheaper and safer.

How to run a fair medical ASR bake-off

  1. 1Assemble your own corpus

    Collect 30 to 50 recordings that match your specialties, accents, devices and room noise. Vendor demo audio is unrepresentative by construction.

  2. 2Write reference transcripts and reference notes

    Have a clinician produce both. Without references you can compare engines to each other but not to correct.

  3. 3Test streaming and batch separately

    They are different products. Record end-to-end latency from your own client, including network, not just model inference time.

  4. 4Score diarization explicitly

    Count speaker-attribution errors on two-party conversation; this is usually what breaks downstream summarisation.

  5. 5Measure clinically weighted errors

    A wrong dosage and a wrong filler word are not the same error. Weight terms whose corruption would change care.

  6. 6Price at real volume

    Model per-minute cost at your realistic and ten-times volumes, including minimums and any charge for streaming or diarization features.

  7. 7Cost the documentation layer honestly

    Add the engineering weeks and ongoing clinical evaluation from the integration effort model before comparing against a scribe API.

Frequently asked questions

Which medical speech-to-text API is best?
Deepgram leads our medical speech-to-text ranking on developer experience, published pricing and streaming maturity. AWS Transcribe Medical and Azure AI Speech are reasonable defaults if you are already committed to that cloud and want HIPAA-eligible service coverage under an existing agreement. See the ranking.
Do medical speech-to-text APIs produce clinical notes?
No. They return transcripts, optionally with speaker labels, timings and confidence values. Turning that into a structured, template-compliant clinical note is a separate layer that you either build or buy. Clinical documentation API guide.
Why do you not publish word error rates for these vendors?
Because no vendor publishes a corpus, audio-condition set and reference-transcript protocol complete enough to reproduce a figure. Publishing unreproducible numbers would make this site less accurate. Measure on your own audio instead — the guide above explains how. Methodology.
Is HIPAA-eligible the same as HIPAA-compliant?
No. A cloud provider listing a service as HIPAA-eligible means the service may be used with PHI under a BAA when configured correctly. Compliance still depends on your configuration, retention settings, access controls and documentation. HIPAA guide.
What is the most accurate medical speech-to-text API?
No vendor publishes a reproducible word error rate on a shared clinical test set, so any ranking of accuracy claims is a ranking of marketing. Deepgram's medical model line, AWS Transcribe Medical and AssemblyAI's medical mode are the credible clinically-tuned options; decide between them by running all three on 20 to 50 of your own recordings with your own scoring script. How we treat accuracy claims.
Is there a HIPAA-compliant speech-to-text API?
Several: AWS, Azure, Deepgram, AssemblyAI and Speechmatics all support healthcare deployments with a BAA available. HIPAA compliance is a property of your configuration and contract, not of the model — the same endpoint is compliant with a BAA, retention limits and PHI-training exclusions, and non-compliant without them. Compliance checklist.
How much does a medical transcription API cost?
Clinically-tuned ASR is priced per minute or per hour of audio, with real-time streaming typically costing more than batch. The per-minute rate is the small number; the durable cost is the documentation layer, evaluation harness and clinical-safety ownership that a transcript alone leaves with you. Cost calculator.
Can I use a general speech-to-text API for clinical notes?
For transcription, often yes — general models handle conversational speech well and custom vocabulary closes much of the terminology gap. For documentation, no: medication names, dosages and negation are exactly where general models fail silently, and a plausible wrong dose is worse than a missing one. When general APIs are enough.
Do medical speech APIs support real-time streaming?
Deepgram, AssemblyAI, Azure AI Speech, Speechmatics and AWS all document streaming transcription over WebSocket-style transports. Streaming roughly doubles client complexity — reconnection, partial results, buffering, mid-encounter failure — so adopt it only if clinicians genuinely need in-encounter output.
Can I run medical speech-to-text on-premises?
Speechmatics is the most straightforward container-deployable option, and the hyperscalers keep processing inside your own cloud account. On-premises is worth its operational cost mainly when a customer contractually refuses external audio processing.
What accuracy metric should I ask a vendor for?
Ask for word error rate with the test set, audio conditions, speaker mix and scoring script disclosed, plus a clinically-weighted error rate for medications, dosages, laterality and negation. Any number without a reproducible method is unusable in your own security or clinical review.

Evidence & sources

Every factual claim on this page traces to one of the primary references below. Each entry records what it supports and its evidence tier, so documentation can be told apart from judgement.

  1. [1]Deepgram API documentation

    Deepgram · Vendor documentation · Tier A — primary documentation

    Supports: Self-serve access, streaming endpoints, model options and documented rate limits.

  2. [2]Deepgram pricing

    Deepgram · Vendor documentation · Tier A — primary documentation

    Supports: Published per-minute pricing used in our commercial transparency scoring.

  3. [3]Amazon Transcribe Medical documentation

    Amazon Web Services · Vendor documentation · Tier A — primary documentation

    Supports: Medical ASR capabilities, streaming support and regional availability.

  4. [4]AWS HIPAA-eligible services and BAA

    Amazon Web Services · Vendor documentation · Tier A — primary documentation

    Supports: Which services are covered by the AWS BAA — the basis of our AWS compliance score.

  5. [5]Azure AI Speech documentation

    Microsoft · Vendor documentation · Tier A — primary documentation

    Supports: Streaming, custom speech models, language coverage and regional deployment.

  6. [6]Microsoft HIPAA / HITECH compliance offering

    Microsoft · Vendor documentation · Tier A — primary documentation

    Supports: BAA coverage for Azure services processing PHI.

  7. [7]AssemblyAI API documentation

    AssemblyAI · Vendor documentation · Tier A — primary documentation

    Supports: Self-serve onboarding, streaming, summarisation and PII redaction features.

  8. [8]Speechmatics documentation

    Speechmatics · Vendor documentation · Tier A — primary documentation

    Supports: Container and on-premise deployment options, language coverage, streaming API.

  9. [9]Word error rate — measurement definition and limitations

    NIST speech recognition evaluation literature · Methodology · Tier B — published methodology

    Supports: Why a headline WER figure without a stated corpus, audio condition and reference-transcript protocol is not comparable across vendors.

  10. [10]Business Associate Contracts — sample provisions

    U.S. Department of Health & Human Services · Regulation · Tier A — primary documentation

    Supports: The contractual clauses a documentation vendor's BAA must contain.

Source tiers are defined on the methodology page. Outbound links are unaffiliated and carry no commercial relationship.

Continue