Solving the accuracy gap that puts patients at risk

The problem with "good enough" transcription

In most software, a transcription error is an inconvenience. In clinical intake, it can be a medical error. A misheard drug name, a garbled condition, a mistranscribed dosage — any one of these can set off a chain of decisions with serious consequences for a patient.

This is the environment Humetrix operates in. The company builds multilingual intelligent intake software for healthcare providers: a voice-enabled platform that lets a nurse, physician, or paramedic conduct a full patient intake interview in the patient’s native language, then delivers a clinically coded summary in the provider’s language. No interpreter required. No language barrier in the room.

The platform supports 27 languages and is designed for exactly the settings where translation is hardest: busy hospitals, mobile healthcare units, telehealth consultations, and mass-care events where staff are stretched and the patient population is anything but uniform.

For a platform built on precision, transcription accuracy is not “just” a feature. Here, getting it wrong means a patient who wasn’t properly heard.

Watch the Humetrix platform in action:

What Whisper couldn’t deliver

Humetrix originally built its transcription layer on OpenAI’s Whisper v3 large model. In general use, Whisper performs well. In clinical use, it didn’t.

Medical terminology exposed the model’s limits. Drug names, anatomical terms, and condition names sit outside the probability distributions that general-purpose models are trained to favour. The error rate on these terms was, in Humetrix’s assessment, unacceptably high for clinical deployment.

The stakes made a switch non-negotiable. Humetrix needed a transcription engine that could handle the full complexity of spoken medical language, across multiple languages, with the reliability that patient safety demands.

Clinical accuracy at scale

After evaluating alternatives, Humetrix moved to Speechmatics. The deciding factor was performance on multilingual medical vocabulary — the specific domain where accuracy matters most, and where the previous solution had failed.

Speechmatics is embedded directly into the Humetrix intake workflow. The platform runs on-premises in AWS elastic containers, with Speechmatics operating in batch mode to process spoken patient answers in real time during the clinical interview. The clinician selects the patient’s country and language at the start of the session; Speechmatics handles transcription; Humetrix’s own medical extraction and coding layer converts the output into structured, EMR-ready clinical data.

The result is a clinician who can conduct a complete intake interview with a patient who speaks no shared language, receive an accurate clinical summary in their own language, and get automated medication equivalence — identifying, for example, the French equivalent of a drug the patient takes at home under a different name.

Proven at the Paris 2024 Olympics

The scale of what Humetrix’s platform can handle was put on record at the Paris 2024 Olympic Games. Healthcare teams used the platform to support medical care for an event that brought together millions of international spectators and thousands of athletes from around the world. The patient population spanned 97 countries and more than 22 languages.

In that environment, high volume, compressed timeframes, maximum linguistic diversity, the platform performed. Clinical staff could triage patients accurately regardless of language, identify medications correctly, and avoid the kind of errors that come from relying on improvised translation or incomplete intake.

That deployment established what the platform can do under pressure. Speechmatics now powers the transcription layer that takes it further, bringing voice-enabled multilingual intake to the same high-stakes clinical settings — with the accuracy that every patient deserves.