Healthcare: transcription of consultations

#Transcribing a consultation locally: what is permitted, and under what conditions

Yes, a doctor can transcribe consultations with AI without the audio leaving their workstation: Whisper transcribes the speech, a local LLM (via Ollama) drafts a structured report, and the professional reviews and approves it. This does not remove the legal obligations. The 2026 HAS and CNIL guide reminds us that introducing an AI system does not change the rules on professional confidentiality under Article L. 1110-4 of the French Public Health Code. You must also inform the patient and respect their right to object. Hosting by a third party requires an HDS-certified host. Local processing specifically avoids this transfer and its risks. This guide builds the pipeline, then details the framework and essential checks while correcting a few common misconceptions.

According to that same guide, automatic consultation transcription is the most common generative AI use case in healthcare. The issue, then, is not whether it can be done, but how to do it properly.

#The French framework: what the official guides say

The table covers each requirement and its concrete effect on a local pipeline. It also corrects two common misconceptions: patient consent is not the legal basis for processing, and HDS is not required by default, but when a third party hosts the data.

#Two misconceptions to abandon

First misconception: “ChatGPT is formally prohibited.” The rules do not target any particular tool. The issue is structural: sending health data to an online service moves it off your device, which calls for a certified hosting provider, a data-processing agreement, and appropriate notice. The HAS and CNIL guide also identifies a risk of violating medical confidentiality when sensitive data is integrated into prompts sent to online conversational agents.

Second misconception: “local means compliant.” Local processing eliminates the transfer, not the obligation to keep a register, inform patients, secure the workstation, and consider whether an impact assessment is required. The guide lists the DPIA among the deployer's obligations, where applicable: get support to decide this point, as well as whether your tool may qualify as a medical device.

#The stack: Whisper, a local LLM, and optional diarization

For throughput, the faster-whisper documentation provides a dated reference point: 13 minutes of audio transcribed in 1 minute 03 seconds with the large-v2 model in fp16, using 4,525 MB of VRAM on an RTX 3070 Ti with 8 GB (maintainer benchmark, CUDA 12.4). The same maintainers measured 59 seconds using 2,926 MB in int8. These are their measurements, on their machine, not ours: they indicate that a modest graphics card can transcribe faster than real time, without promising the same results on your system.

#Installation

python3 -m venv venv
source venv/bin/activate

pip install faster-whisper ollama
# Optionnel pour la diarisation :
pip install pyannote.audio

# Le modèle Whisper se télécharge au premier usage.
# Modèle de rédaction via Ollama :
ollama pull qwen3.5:9b

With a NVIDIA GPU, faster-whisper requires the cuBLAS and cuDNN 9 libraries for CUDA 12. Without them, the program fails while loading the model: install them before testing, or switch to CPU mode.

#Transcribe audio

The script below sets the language, enables the silence filter (VAD), and timestamps each segment. Adapt the first line: device cuda with float16 for a NVIDIA GPU, device cpu with int8 for a Mac or a PC without a compatible card.

from faster_whisper import WhisperModel

# GPU NVIDIA : device="cuda", compute_type="float16"
# Sans GPU compatible : device="cpu", compute_type="int8"
model = WhisperModel("large-v3", device="cuda", compute_type="float16")

MOTS_CLES = "amoxicilline, paracétamol, HbA1c, tension artérielle"

def transcribe(audio_path):
    segments, info = model.transcribe(
        audio_path,
        language="fr",
        beam_size=5,
        vad_filter=True,        # coupe les silences
        hotwords=MOTS_CLES,     # vocabulaire attendu
        word_timestamps=False,
    )
    lines = []
    for s in segments:
        ts = f"[{int(s.start // 60):02d}:{int(s.start % 60):02d}]"
        lines.append(f"{ts} {s.text.strip()}")
    return "\n".join(lines)

if __name__ == "__main__":
    import sys
    print(transcribe(sys.argv[1]))

The hotwords parameter, exposed by faster-whisper’s transcription function, lets you suggest expected terms to the model: drug names, acronyms, and tests. It helps but does not guarantee accuracy: always review proper names and dosages.

#Whisper can write what nobody said

The large-v3 model card warns that predictions may contain text not present in the audio, which it calls hallucination. During a consultation, this can produce a plausible sentence invented during silence, a modified number, or a medication name that sounds similar but is wrong. The VAD filter reduces the risk during silence; it does not replace proofreading.

#Write the structured report

The second script sends the transcript to a local model through the Ollama API. Two settings matter: a low temperature to limit hallucinations, and a sufficient context window. Ollama’s documentation indicates a default window of approximately 4,000 tokens under 24 GiB of VRAM. A twenty-minute consultation represents around 3,000 words, or several thousand tokens (an estimate that varies with speaking rate): without an explicit num_ctx, the end of the transcript may be truncated. The example therefore sets 16,384 tokens.

import requests

SYSTEM_CR = """Tu es un assistant pour médecin généraliste français.
À partir d'une TRANSCRIPTION BRUTE d'une consultation, tu produis un
compte-rendu médical structuré.

FORMAT DE SORTIE :
## Motif de consultation
## Antécédents (mentionnés)
## Examen clinique
## Diagnostic / Hypothèses
## Traitement / Prescriptions
## Suivi

RÈGLES :
- Reste TEXTUEL : n'ajoute aucun élément non mentionné dans la transcription.
- Reformule en vocabulaire médical standard.
- Si une section n'est pas abordée, écris "Non mentionné".
- Ne complète jamais une posologie ou un diagnostic absent.
"""

def summarize(transcription):
    r = requests.post("http://localhost:11434/api/chat", json={
        "model": "qwen3.5:9b",
        "messages": [
            {"role": "system", "content": SYSTEM_CR},
            {"role": "user",   "content": transcription},
        ],
        "stream": False,
        "options": {"temperature": 0.2, "num_ctx": 16384},
    })
    return r.json()["message"]["content"]

This prompt applies the principles from the guide to system prompts: verifiable rules, an explicit output format, and defined behavior when information is missing. The instruction “Not mentioned” matters: it prevents the model from filling an empty section with a plausible assumption.

#Diarization: distinguishing doctor and patient

Diarization assigns each segment to a speaker (SPEAKER_00, SPEAKER_01). It is mainly useful when you record the consultation itself, with two speakers; for a dictation by the doctor alone, it is unnecessary. The pyannote repository currently recommends the community-1 pipeline: you must accept its terms of use on Hugging Face and create an access token to download it. Inference then runs locally.

import torch
from pyannote.audio import Pipeline

pipeline = Pipeline.from_pretrained(
    "pyannote/speaker-diarization-community-1",
    token="HUGGINGFACE_ACCESS_TOKEN",
)
pipeline.to(torch.device("cuda"))  # si un GPU est disponible

output = pipeline("consultation.wav")
for turn, speaker in output.speaker_diarization:
    print(f"{turn.start:.1f}s - {turn.end:.1f}s : {speaker}")

#Typical law-firm workflow

1. 01
   Inform the Patient

   Before recording, state that the consultation or dictation will be transcribed by a local tool and that the patient may object. If you record their voice, obtain their consent. Document the information in the record.
2. 02
   Save

   Two options: dictate 1 to 2 minutes of notes after the consultation (the patient is no longer being recorded), or record the conversation (verbal consent required, diarization useful). Dictation minimizes the data collected.
3. 03
   Place the file on the target machine

   Transfer the audio (WAV, MP3, M4A) to a dedicated, encrypted folder. A small script monitors the folder and triggers processing.
4. 04
   Run the pipeline

   The script transcribes, summarizes, saves the report as text next to the audio, and logs the operation without writing any medical content to it.
5. 05
   Review and validate it yourself

   The doctor who saw the patient reviews the report, corrects it, and then enters it into their practice-management software. A third party who was absent from the consultation cannot spot hallucinations.
6. 06
   Delete the raw audio

   Once the report is approved, delete the audio and intermediate transcript: the data no longer serves a purpose, and the report becomes part of the medical record.

#What to check during proofreading

#Retention, security, and traceability

Record of processing activities

The CNIL and French Medical Council guide requires you to keep a record of processing activities; add a line describing this local pipeline and its purposes.

Retention period

The report follows the medical record's retention period (20 years after the last consultation, according to the professional body's recommendation). The raw audio, no longer needed once the report is validated, is deleted.

Encryption and backups

Encrypt the disk (FileVault, BitLocker, LUKS) and back it up regularly to encrypted media stored outside the office. A stolen workstation with unencrypted health data is a data breach.

Traceability

Keep a record of the patient's information, the tool versions, and the physician's validation: these are your evidence in the event of an audit.

• Medical transcription and local LLMs: patient data compliance
• Whisper + Ollama locally: 100% offline transcription
• Medical record summarization
• Encrypt the model drive
• WhisperX: word-level timestamps and speakers
• Master system prompts

• Source: HAS and CNIL guide on the proper use of AI systems in healthcare settings
• Source: CNOM and CNIL guide on protecting patient data
• Source: faster-whisper
• Source: the Whisper large-v3 model card
• Source: pyannote-audio