How Autoscriber gets medical terminology and dosages right in patient consultations using Reson8

Autoscriber turns clinician-patient consultations into clinical notes, structured data elements, and decision support. And these healthtech products are only trustworthy when the dosages and readings in them are accurate. After testing several speech-to-text providers, Autoscriber moved its transcription to Reson8. Dutch, German, French and English consultations now come back more accurate, five times faster, and on GPUs owned by Reson8 in the Netherlands.

At a glance

One misheard number ends up in the patient record

Autoscriber records the consultation and drafts the note for a clinician to sign off on. Notes contain a range of medical terminology, like medication names, dosages, lab values and readings. A transcription error on any of those could lead to an incorrect patient record and a potential patient safety issue. If a clinician catches an error, they stop trusting the scribe.

General speech engines struggle on numbers because they're short, carry little acoustic context, and can be written differently between languages. A body temperature of ‘37.8 degrees’ might use a decimal in English, but needs a comma separator (‘37,8’) in Dutch. A transcription engine has to get both the digits and the format right, which is where Reson8 stood out.

“In a consultation, getting the numbers wrong can be very dangerous for patients. We tested several providers on our own recordings, and Reson8 was the first one where dosages and alphanumerics consistently came back correctly.”Rik Matena, CPO, Autoscriber

Autoscriber tested on its own consultations

Autoscriber benchmarked Reson8 on real consultation audio rather than a publicly available test set, evaluating hardest where previous engines had fallen short.

Medical accuracy: On Autoscriber's consultation benchmark, Reson8's base model with no medical customization beat a dedicated medical model from another provider.

Mixed languages: Some consultations move between languages. On a mixed Dutch and German test set, word error rate (WER) fell by 62% compared to Autoscriber's previous provider when switching to Reson8. On prerecorded audio, automatic language detection picked the right language reliably.

Speed: Final transcripts came back five times faster than on Autoscriber's previous setup.

Fully EU sovereign speech AI that runs on hardware Reson8 owns

Under GDPR, consultation audio is considered health data, so Autoscriber's hospital customers have a privacy officer who approves every company that processes it. Because Reson8 runs on GPUs it owns and runs in the Netherlands, there is no reliance on a cloud provider and no third party handles the audio on Reson8's behalf. No audio is stored after transcription, and no sub-processor sits in the audio path.

That gives Autoscriber a short answer for hospital customers, which is that audio is processed in the Netherlands by one Dutch company, and nothing is kept.

“Reson8 gave us two things we couldn't get elsewhere: far fewer errors on the clinical details that matter, and a data story hospitals immediately trust. Their own hardware in the Netherlands, no subprocessors, zero retention.”Danny van Rijn, COO, Autoscriber

Running medical or multilingual transcription? Benchmark your consultation audio through Reson8 and compare accuracy against your current engine. Get in touch, get started at console.reson8.dev, or read our trust center.