Voice Input: Dictation
9 min readUpdated on 12 November 2026
⚠️ Rewritten on 2026-08-17 based on the actual screen. The previous version described a panel that never existed: a
Extract fieldsbutton, a reliability score for each field with green, yellow, and red colors, three buttons per row (Accept/Edit/Discard), and aSettings > Dictationsection with silence threshold and transcription retention. None of this is in the product. What follows is.
Dictation transcribes what you say and, where the form allows it, suggests pre-filled fields. It never writes directly to the medical record: between your voice and the saved data, there is always a review and a button pressed by you.
Where to dictate
The button appears in three places, with a different label in each because "dictate what" changes depending on the location:
Dictate medical history, in theOverviewsection of the record,Medical historytab;Dictate the session, in the treatment module, next to the notes;- in the clinical assessment, when the practice has that module active.
At rest, it is a single line: a button and a phrase. The box appears when there is content inside.
Prerequisites
- Account with
physicianrole and clinical access to the patient. - Working microphone with browser permission granted. Transcription quality depends more on ambient noise than on the microphone.
- Connection: transcription occurs on a service, not in the browser.
Step 1: Dictate
Press the button. A red dot appears with the text Listening, and below, in Transcription, the text appears as you speak: «Speak freely: the text appears here as you speak».
Two buttons: Done closes the dictation and moves to review, Cancel discards it.
Step 2: Review
At the end, the transcribed text appears in an editable area, under a notice worth reading once:
Review before using it. Automatic transcription is most likely to make mistakes with drugs, dosages, and technical terms: correct them below.
If the form allows field extraction, next to the notice appears the extraction reliability as a percentage. This is a single number for the entire extraction, not per field, and it is a technical indicator: it shows how clear the model found the text, not how correct what you said is.
Step 3: What to do with it
Three buttons, each doing different things:
Discard: deletes the transcription.Use text: takes the text as is and places it in the target field (for example, at the end of the session notes). It only appears where that text has a destination: elsewhere, it would be a button that deletes for no reason, and it has been removed.- the application button (
Suggest for the recordin medical history,Fill fieldsin treatment): takes the recognized fields and transfers them to the form, where they remain editable. It only appears if the extraction produced something.
⚠️ Even after applying the fields, saving is a separate action. Applying fills the form; what goes into the record is what you save.
Uploading a recording made elsewhere
If the visit was recorded with a recorder or a phone, next to Dictate there is Upload a recording. First you tick «The patient consented to the recording of the visit», then you choose the file:
- MP3, WAV or M4A audio: it is transcribed like a dictation. A long recording is sent in several parts and the screen says which part it has reached;
- text already transcribed in TXT, Markdown or SRT (subtitles: numbers and times are removed automatically);
- a DOCX or a PDF cannot be read: export it to TXT.
Then everything works like dictation: the transcript stays in the visit, with the file name and the date on which you declared the consent, and the extracted fields are reviewed before saving. The audio is not kept. Files over 400 MB cannot be uploaded; an M4A over 20 MB must be exported to MP3 or WAV.
What dictation fills in, and what it doesn’t
This is where expectations most often break, so it’s better to measure than to promise.
In treatment, product, quantity, lot, off-label use, and its justification are suggested. Not filled in are device parameters (wavelength, fluence, spot, frequency, pulse duration, passes, cooling, endpoint), nor dilution, UDI, or lot expiration: these must be entered manually.
Dictated areas do not become dots on the map. They are added at the end of the notes in the form [dictated areas: …], along with any [suggested category: …], because marking an area requires its exact code. To transfer them to the map, use the Auto-extract areas from text button: see Treated Areas.
⚠️ Dictation is in Italian. Even with the interface in English, recognition and extraction work in Italian.
Why you choose the scale grades
This is the limitation that looks like a defect and isn't, so here it is in full.
Forms contain two different kinds of field, and dictation treats them differently on purpose:
- facts reported by the patient: allergies, medications, conditions, pregnancy, smoking, skincare, past treatments. Here "she is not pregnant" →
nois transcription: dictation simply fills them in; - graded judgements on validated scales: Fitzpatrick, Glogau, and the absent / mild / moderate / marked steps of the physical examination. Here assigning the grade is a medical act: it is you who, looking and touching, decide which step this patient is on.
If the software derived the grade from your description, from "the skin gives way a lot when I lift it" to "firmness: marked", it would not be transcribing: it would be interpreting clinical information to support a decision. The European Commission guidance on software qualification (MDCG 2019-11 (opens in a new tab), §3.1) states that software which "process, analyse, interpret, calculate, create or modify medical information" may qualify as a medical device when it does so "so that it serves as a decision support or suggests an action to be taken by the user".
Two clarifications in the same guidance settle it:
- the risk of harm is not a criterion for qualification. So "the doctor confirms it anyway" does not remove the qualification: ticking a box next to an already-written grade is not enough;
- natural language processing is not exempt as "simple search" when it contributes to a medical purpose.
Under Rule 11 of the MDR this would mean class IIa at minimum: notified body, clinical evaluation, CE marking. Fibonacci is a medical record, not a medical device, and this line is what keeps the distinction.
How it works in practice
There is only one boundary, from "gives way a lot" to "marked" is your call, and dictation does everything else. Every scale field ends up in one of three states:
| If you… | Dictation… |
|---|---|
| said the grade ("firmness moderate") | fills it in, and shows your own words underneath: "You said: …" |
| only described it ("gives way a lot when I lift it") | does not grade. It puts your sentence next to the field, under "For you to grade", and opens the selector empty: grading is one tap |
| did not mention it | leaves the field as it was |
You lose nothing in the second case either: the description still goes into the section's free-text fields (inspection, palpation, observed signs) and into the transcript.
What you gain
- The record says what you said. Next to every filled grade the sentence it came from remains, so on re-reading (or if someone else reads the record) it is clear the judgement is yours and where it comes from.
- No plausible grade that isn't yours. An already-written value gets confirmed out of inertia at the end of the day; an empty box does not. That is why the inferred grade is not even offered to you as a suggestion to tick.
If you dictate a grade and the field does not fill in even though you said it, that is not the intended behaviour: it is a defect, and it should be reported with the exact sentence you spoke.
Clinical responsibility
The principle is non-negotiable: the system does not write anything to the record without an explicit action by the physician. Every transcribed text and every suggested field require review and an affirmative action. Responsibility for correct completion remains with the person who signs the record.
Audio flow privacy
Audio is sent to Mistral AI SAS (Paris, France), which provides the artificial intelligence models used by Fibonacci (dictation transcription, extraction of fields from dictated texts, assistant) through its European endpoint. Training of models on the data sent is switched off in Fibonacci's organisation. Zero retention of content is not yet active and the provider has not yet confirmed it in writing: for this reason dictation is not used with real patients. The provider's data processing agreement: https://legal.mistral.ai/terms/data-processing-addendum (opens in a new tab) .
If you prefer not to use dictation for a visit, complete the form manually: no audio trace remains anywhere.
Tips
- Speak at a natural pace, without over-enunciating: the model is calibrated for spontaneous Italian speech, and slowing down worsens the result.
- No voice commands like «period» or «new line»: punctuation is added automatically.
- Drugs in full, active ingredient and dose: «pantoprazole forty milligrams one tablet in the morning».
- One voice at a time. If the patient speaks at the same time as you, transcription quality drops.
- Always double-check numbers. Dosages and lots are exactly what transcription gets wrong most often, and they are also what matters most.
Troubleshooting
The microphone is not detected. Check browser permissions (in Chrome, the padlock to the left of the address, Microphone entry) and operating system settings: a microphone turned off at the system level is not accessible to the browser.
A red error appears below the button. The message states the cause: usually, it’s denied permission or the transcription service being unreachable.
The transcription arrives, but no fields are suggested. The application button only appears if the extraction recognized something. You can still use Use text and correct manually.
I dictated the areas, but the map is empty. This is the expected behavior: see above, «What dictation fills in, and what it doesn’t».
See also
- Patient record creation and management
- Treated areas: on the photo and 3D model
- Recording a treatment
- Audit log and access traceability
Last revision: 17 agosto 2026