Medical dictation: your reports dictated, your terms spelt right

A dictated report is only worth something if every term in it is right. General-purpose voice dictation recognises everyday conversation very well, but it stumbles over the very stuff of medicine: names of conditions, surgical techniques, eponyms, devices, drugs. TechVoice was designed for this precise problem, starting with orthopaedic surgery. It works in French and in English.

Why general-purpose dictation struggles with medical vocabulary

At every moment, a speech recognition model picks the most likely words, based on everything it heard during training. Medical terms are rare in that material, so when the sound is ambiguous, the everyday word wins. In French, dyspnée (dyspnoea) becomes dispersée (dispersed), and “hallux valgus” becomes “Puralux valgus”. A report in which this happens is unusable, even if everything else is perfect.

In our orthopaedic trials, a transcription model used on its own missed up to one specialist term in three. It is the time spent fixing these errors, document after document, that makes people give up on dictation.

How TechVoice learns your specialty

On top of speech recognition, TechVoice adds a correction step that knows your vocabulary, fed by three sources:

1. Your specialty. Entered when you sign up, it is used to look up its terms in the scientific literature of the HAL open archive, in the language you dictate in.
2. Your sample documents. Upload a few reports or letters (PDF, Word): TechVoice extracts the terms, then forgets the text.
3. Your corrections. Every word you fix becomes a “misheard → correct” pair, applied to your next dictations.

On our French test corpus, the text displayed during dictation gets about 25% of specialist terms wrong; after correction with the practitioner’s vocabulary, about 5%. These figures come from synthetic voices and are optimistic, but the order of magnitude of the effect holds from one measurement to the next. The details are in our article on how AI voice dictation works.

Reports that stay true to what you dictate

Your name goes on every report. TechVoice’s correction only touches misheard words: it adds nothing, removes nothing and rewords nothing. An automatic check compares the corrected text with the recognised text, word by word and in order; if too much has changed, the raw transcript is kept. Moving a negation is enough to reverse a diagnosis: “no pulmonary embolism” is not “pulmonary embolism”.

Every correction stays visible: one click shows what was changed. You always review a text before using it, and you remain its author.

Documents you can dictate

• Consultation notes, dictated straight after the patient leaves, while everything is fresh in your mind.
• Operative reports, with techniques, surgical approaches and devices.
• Letters to colleagues: one-click formatting turns your dictation into a letter or an email, ready for you to review.
• Follow-up notes and summaries, filed in folders and easy to find again with search.

You can also upload a recording made elsewhere, such as a voice memo between two consultations: TechVoice transcribes and corrects it in the same way, up to 60 minutes long.

What about health data?

TechVoice does not hold HDS certification, France’s certification for hosting health data. So do not dictate anything that could identify a patient: no name, no date of birth, no record number. The simplest way to work: dictate the medical content in TechVoice, then add the patient’s identity in your medical records software.

Your voice and your texts are transcribed and corrected by Mistral AI and hosted by Scaleway, in France. Audio is not kept after transcription. Full details are on our Privacy page (in French).

Try it on your own reports

The most reliable way to judge is still to try it on a week of real documents. You can start now: you get free points when you sign up, with no credit card required. To estimate how much time you could save, see our time-savings calculation, document by document.