Radiology speech recognition: dictation accuracy for diagnostic workflows

A 2024 study in Radiology: Artificial Intelligence found clinically significant speech recognition errors in 3.2% of 3,233 radiology reports reviewed, with longer reports, resident dictation, and overnight shifts all associated with higher error rates.

These errors include wrong words and missing negations that can alter diagnoses; measurement substitutions can do the same. If you're building clinical documentation tools or radiology reporting products, your STT API choice directly affects risk in the final report.

Most content on this topic covers desktop dictation software for radiologists. At the API layer, radiology speech recognition depends on model accuracy, vocabulary control, streaming behavior, and compliance architecture. Deepgram's speech-to-text platform addresses those API requirements for clinical documentation workflows.

• The same 2024 study found errors of any kind in 44.2% of radiology reports reviewed, with clinically significant errors being the subset most likely to affect diagnosis.
• Deepgram reports a 6.79% keyword error rate on medical terminology for Nova-3 Medical.
• HIPAA requires a BAA before any cloud STT API processes radiology audio.
• Streaming dictation matters most, but your back end still has to handle asynchronous transport.

General-purpose STT misses too many clinically meaningful radiology terms for diagnostic workflows. You need a model that handles specialized vocabulary and report structure under real dictation conditions.

Radiology reports are packed with specialized terms that general STT models rarely encounter in training data. Drug names and contrast agents create a dense vocabulary problem alongside anatomical structures.

A 2024 study in Diagnostics customizing Whisper Large-v2 for French radiology still produced a 17.12% word error rate after training for that setting.

Wrong-word substitution is the single most common error type, reaching 0.18 errors per report in a peer-reviewed analysis. Gadolinium formulations, iodinated contrast agent names, and subspecialty anatomical terms consistently trip up models trained on general-purpose audio.

Rigid structural conventions govern every radiology report: findings, impressions, recommendations, and comparison sections have to land in the right place. General-purpose models don't handle these patterns well, and can break template field mappings and introduce errors at section boundaries.

Longer, complex reports carry a disproportionate error burden, with reports exceeding 25 sentences averaging 1.23 errors, while shorter reports showed lower rates. Cross-sectional reports such as CT and MRI also showed higher error rates than plain radiography, a pattern confirmed at scale by a Mayo Clinic analysis of 213,977 reports, which reported an odds ratio of 3.72.

Reading rooms and remote dictation setups introduce background noise and device variability that pull accuracy down. When masks were worn, dictation errors rose from roughly 21.7 to 27.1 errors per 1,000 words.

The Nova-3 Medical announcement describes performance with audio captured from iPads, laptops, and phones, with background chatter and equipment sounds present, conditions most radiologists actually work in.

Low WER, reliable keyword handling, and structured-workflow support are the minimum requirements, along with streaming behavior that fits how radiologists actually dictate.

Any meaningful STT evaluation for this use case should include keyword error rate, which tracks accuracy on the clinical terms that carry diagnostic weight.

Public radiology benchmarking is still hard to compare because some datasets are private and not peer reviewed, so treat third-party benchmark claims carefully and rely on testing with your own audio.

Your front end should support streaming dictation, while your back end tolerates asynchronous message and document transport. One architecture for this pattern records audio through a web interface, processes it on a remote server, and displays transcriptions in real time.

Batch processing still exists as a legacy pattern from the era of human transcription services, but it shouldn't define your integration. In practice, your STT integration should support real-time streaming at the dictation UI layer while also handling asynchronous transport, such as DICOM SR objects and HL7 messages, on the back end.

Subspecialty vocabulary changes fast enough that a static medical model may not be enough on its own, and request-level control over preferred terminology is what separates a working integration from one that drifts over time.

Neuroradiology, musculoskeletal imaging, breast imaging, and chest radiology each carry distinct vocabulary sets, and that variation has measurable consequences: the same Mayo Clinic analysis that identified the OR 3.72 modality finding also confirmed that error rates varied significantly by imaging subspecialty.

A one-time model configuration may not cover that variation well enough. Request-level vocabulary injection lets you push subspecialty-specific terms into the model at inference time. Keyterm Prompting addresses this by accepting up to 100 custom terms per API request.

If your STT stack processes radiology audio, it handles protected health information, which pulls deployment architecture and vendor contracts into the product decision.

HHS guidance is explicit: when a covered entity uses a cloud STT API to process audio containing ePHI, the API provider is a business associate. A BAA must be in place before any PHI processing begins, and that covers more than you might expect: both the audio stream and the resulting transcript qualify as ePHI.

Keep in mind that a cloud provider is still a business associate even if it processes or stores only encrypted ePHI and lacks the decryption key, and HHS has entered resolution agreements with covered entities that stored protected data on cloud servers without that agreement in place.

Your BAA must also address the subcontractor chain: the STT vendor's GPU infrastructure providers, storage services, and other subprocessors all need coverage.

Different deployment models change who controls PHI, who needs a BAA, and how much operational burden you carry. Geographic location of PHI processing becomes a risk analysis documentation requirement under 45 CFR § 164.308(a)(1)(ii)(A).

For organizations that need full control over data residency, on-premises deployment keeps audio and transcripts within the hospital's security boundary. No external BAA is required for the STT component in this model.

VPC deployment offers a middle ground. It provides network isolation with region-configurable data residency while reducing the burden of managing GPU infrastructure. Deepgram maintains HIPAA-aligned deployments with BAA terms handled through sales and enterprise agreements. Nova-3 Medical supports managed cloud, on-premises, and VPC deployment options.

Radiology STT output has to land in the reporting systems and imaging workflow your users already depend on. A 2023 study in Insights into Imaging describes SR² as an explicit integration pattern, where radiology speech recognition populates structured report fields. DICOM SR objects carry measurements and observations from modalities.

HL7 v2 messaging remains the predominant mechanism for RIS integration in deployed systems. HL7 FHIR DiagnosticReport resources represent the forward direction. FHIRcast provides context synchronization across applications, keeping the PACS viewer and reporting application aligned on the same study. Your reporting application needs that synchronization before STT output can be correctly attributed.

For radiology workflows, the main questions are terminology handling, request-level vocabulary control, and deployment flexibility.

According to the Nova-3 Medical announcement, Nova-3 Medical achieved a 6.79% keyword error rate on medical terminology and a keyword recall rate of 93.99%.

The benchmark used a diverse medical audio dataset designed to reflect real-world clinical scenarios, incorporating drug names and procedure terminology from both public datasets and proprietary customer audio.

For radiology dictation, that maps to the high-risk vocabulary categories you care about most: negation words like "no," measurement units, laterality terms, and contrast agent names.

With up to 100 custom terms injectable per API request, Keyterm Prompting lets you tailor the model to each subspecialty without retraining. For a neuroradiology workflow, you can push specific contrast formulations and anatomical structures; for breast imaging, you can swap in mammography-specific vocabulary instead.

Terms also update dynamically, so as new drugs, procedures, or diagnostic terms emerge, you can add them to your prompt list immediately. That flexibility directly addresses the subspecialty vocabulary variation the Mayo Clinic study identified.

Healthcare teams usually need deployment choice as much as they need accuracy. Nova-3 Medical supports managed cloud, on-premises, and VPC deployment. The model announcement describes data encryption at rest and in transit plus access controls.

It also describes continuous monitoring. For hospital clients running security reviews, on-premises deployment removes the need for an external BAA on the STT component itself. VPC deployment lets you pin processing to specific cloud regions while maintaining network isolation. Managed cloud works for organizations whose risk analysis supports it under an executed BAA.

Before shipping radiology-grade STT, test it against your own audio, workflows, and compliance requirements.

Start testing Nova-3 Medical against your radiology audio samples. Build a Keyterm Prompting list with your target subspecialty's highest-frequency terms. Include contrast agents and anatomical structures, along with procedure names.

Test negation handling explicitly by dictating sentences with "no evidence of" and "unremarkable" patterns. These are the error categories that carry the most clinical risk.

The managed cloud API is the right place for prototyping and accuracy validation. As you move toward hospital deployments, work with Deepgram's sales team to scope the right deployment model and execute BAA terms for your architecture. Pricing scales with usage. See current rates at Deepgram pricing.

You can test Nova-3 Medical against your radiology dictation samples with $200 in free credits when you start for free.

Production WER depends on subspecialty, audio quality, and vocabulary density. Nova-3 Medical's medical benchmark results reflect a diverse clinical audio dataset. Results will vary based on microphone placement and ambient noise. Test with your actual clinical audio before committing to a model.

Abbreviations like "CT" and "MRI" are common enough for most models. The harder problem is measurement units. "mm" versus "cm" substitutions carry staging implications. Keyterm Prompting reduces these errors by biasing the model toward your expected measurement vocabulary at inference time.

You send a list of custom terms alongside your audio. The model uses in-context learning to weight those terms during transcription. You can swap term lists dynamically per subspecialty without retraining or downtime.

Beyond the BAA, your STT vendor must implement audit controls under 45 CFR § 164.312(b), transmission security, and access controls. A December 2024 HHS proposed rule would convert encryption from addressable to required. Architecting for encryption remains prudent regardless of timing.

General-purpose models can post very different error rates across radiology tasks, but the methodologies are often hard to compare directly. Some datasets aren't independently reproducible. Nova-3 Medical uses a different benchmark methodology, so direct comparison is limited. Benchmark your specific audio against each model.