Skip to content
supermax01Public

About

Open-source clinical scribe that runs on your own computer. Turns a doctor–patient visit into a quote-backed note and FHIR records; the clinician approves every fact.

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Latest commit

 

History

4 Commits

Folders and files

Repository files navigation

Auris

Auris

An open-source clinical scribe that runs on your own computer.
It listens to a doctor–patient visit, drafts the note, and lets the clinician decide what becomes part of the record.

License: Apache 2.0 Python 3.12 Status: pre-alpha Runs offline

A fictional visit in Auris, from the day's schedule to the FHIR export, sped up
A fictional visit from the day's schedule to the FHIR export, sped up. Watch it in full quality (MP4).

Important

Auris is early research software. It is not a medical device and is not approved for clinical use. A clinician must review everything it produces. All patients, visits and transcripts in this repository are fictional.

Why Auris?

AI scribes save clinicians hours of typing, but many of them send the conversation to a cloud service and hand back a finished note that is hard to check. Auris takes a different approach:

  • Private by design. Speech recognition, the language model and every check run locally. Nothing leaves the computer unless you export it.
  • Every fact shows its source. Each item in the note links to the exact words in the transcript, with the time and the speaker.
  • Nothing is invented. If a dose or a frequency was not said, the field stays empty and is flagged. The model never writes medical codes.
  • The clinician decides. Nothing becomes a record until a person approves it, and every approval is logged.
  • Standard output. Approved items are exported as FHIR R4 records, the format hospital systems use to exchange data, checked against the US Core profiles with the official HL7 validator.

Features

  • Speech to text with speaker separation, using Whisper locally
  • Extraction of symptoms, conditions, vital signs, medicines and allergies, using a local model through llama.cpp
  • Automatic checks on every fact: the quote must match the transcript word for word, a nearby "no" or "not" cannot be dropped from a quote, and numbers must actually have been said
  • A visit note in the familiar SOAP layout (Subjective, Objective, Assessment, Plan)
  • Medical coding with RxNorm (medicines) and ICD-10-CM (conditions), done by lookup rather than by the model
  • A desktop app for the clinic day: schedule, consent prompt, background note writing, review and export
  • A browser review page and a command-line tool for working with individual visits

A walk through a visit

See the day. Each visit shows what it needs next.

Today's visits

Ask before listening. Auris suggests what to say to get the patient's consent and records the time it was given.

Consent prompt

Let it write. The note is drafted in the background while you see the next patient.

Writing the note

Review. Approve, correct or dismiss each fact, with its quote and the surrounding conversation in view.

Reviewing a fact

Export. Approved records are saved as a FHIR file, together with the visit note. Items that should not become records, such as a symptom or a denied allergy, stay in the note.

Export

From conversation to FHIR R4

Here is part of the fictional demo visit and what Auris exported from it after the clinician approved each fact.

Patient (1:57): I am allergic to sulfa; it gave me a rash years ago. I cannot remember the name on the bottle, but I remember the rash.

Clinician (2:36): Your blood pressure is 126 over 78. You can relax your arm now.

Clinician (3:39): Start acetaminophen 500 milligrams by mouth every eight hours as needed. Let me write that down so you have the instructions at home.

What was said FHIR R4 resource What it contains
"allergic to sulfa; it gave me a rash" AllergyIntolerance (US Core) the substance as spoken ("sulfa"), the reaction ("rash"), marked unconfirmed until a clinician confirms it
"blood pressure is 126 over 78" Observation (US Core Blood Pressure) LOINC 85354-9, with systolic 126 mm[Hg] and diastolic 78 mm[Hg] as separate components
"Start acetaminophen 500 milligrams…" MedicationRequest (US Core) RxNorm 161 (acetaminophen), the dosage exactly as said, and status draft: Auris never signs or sends an order

The export is a FHIR Bundle of type collection. For this visit it holds 3 Condition, 2 Observation, 1 AllergyIntolerance and 1 MedicationRequest resources, plus the Patient, Practitioner and Encounter they refer to. Things that are not records, such as symptoms, denials and family history, stay in the visit note.

Observation: blood pressure 126/78
{
  "category": [
    {
      "coding": [
        {
          "code": "vital-signs",
          "system": "http://terminology.hl7.org/CodeSystem/observation-category"
        }
      ]
    }
  ],
  "code": {
    "coding": [
      {
        "code": "85354-9",
        "system": "http://loinc.org"
      }
    ]
  },
  "component": [
    {
      "code": {
        "coding": [
          {
            "code": "8480-6",
            "system": "http://loinc.org"
          }
        ]
      },
      "valueQuantity": {
        "code": "mm[Hg]",
        "system": "http://unitsofmeasure.org",
        "unit": "mmHg",
        "value": 126
      }
    },
    {
      "code": {
        "coding": [
          {
            "code": "8462-4",
            "system": "http://loinc.org"
          }
        ]
      },
      "valueQuantity": {
        "code": "mm[Hg]",
        "system": "http://unitsofmeasure.org",
        "unit": "mmHg",
        "value": 78
      }
    }
  ],
  "effectiveDateTime": "2026-11-16T08:30:00-05:00",
  "encounter": {
    "reference": "urn:uuid:19b3e431-c0c9-5985-82e8-7e63847e12f8"
  },
  "id": "805b73c6-5cc6-5706-85c9-ddd3931c31c4",
  "meta": {
    "profile": [
      "http://hl7.org/fhir/us/core/StructureDefinition/us-core-blood-pressure"
    ]
  },
  "resourceType": "Observation",
  "status": "preliminary",
  "subject": {
    "reference": "urn:uuid:7f7cae25-8a11-5e8b-9c64-304be9a70b46"
  }
}
AllergyIntolerance: sulfa, reaction rash
{
  "clinicalStatus": {
    "coding": [
      {
        "code": "active",
        "system": "http://terminology.hl7.org/CodeSystem/allergyintolerance-clinical"
      }
    ]
  },
  "code": {
    "text": "sulfa"
  },
  "encounter": {
    "reference": "urn:uuid:19b3e431-c0c9-5985-82e8-7e63847e12f8"
  },
  "id": "88d69937-7b0f-5561-812b-004f45adf7a3",
  "meta": {
    "profile": [
      "http://hl7.org/fhir/us/core/StructureDefinition/us-core-allergyintolerance"
    ]
  },
  "patient": {
    "reference": "urn:uuid:7f7cae25-8a11-5e8b-9c64-304be9a70b46"
  },
  "reaction": [
    {
      "manifestation": [
        {
          "text": "rash"
        }
      ]
    }
  ],
  "resourceType": "AllergyIntolerance",
  "verificationStatus": {
    "coding": [
      {
        "code": "unconfirmed",
        "system": "http://terminology.hl7.org/CodeSystem/allergyintolerance-verification"
      }
    ]
  }
}
MedicationRequest: acetaminophen 500 mg, draft
{
  "authoredOn": "2026-11-16T08:30:00-05:00",
  "dosageInstruction": [
    {
      "text": "500 milligrams by mouth every eight hours as needed"
    }
  ],
  "encounter": {
    "reference": "urn:uuid:19b3e431-c0c9-5985-82e8-7e63847e12f8"
  },
  "id": "5a7a6415-e0d0-5373-8f89-82042562f681",
  "intent": "proposal",
  "medicationCodeableConcept": {
    "coding": [
      {
        "code": "161",
        "display": "acetaminophen",
        "system": "http://www.nlm.nih.gov/research/umls/rxnorm"
      }
    ],
    "text": "acetaminophen"
  },
  "meta": {
    "profile": [
      "http://hl7.org/fhir/us/core/StructureDefinition/us-core-medicationrequest"
    ]
  },
  "reportedBoolean": false,
  "requester": {
    "reference": "urn:uuid:192f4b09-66eb-5d03-a351-dd8286a7164d"
  },
  "resourceType": "MedicationRequest",
  "status": "draft",
  "subject": {
    "reference": "urn:uuid:7f7cae25-8a11-5e8b-9c64-304be9a70b46"
  }
}

The visit note

The same visit also produces a short note in the SOAP layout. Every line says who said it, when, and whether it went into the export file. An excerpt:

## Objective

- Blood pressure 126/78 mmHg (clinician, 2:36) — in the export file · approved by Dr Tobin Hale (clinician) · checked against US Core
- Heart rate 84 /min (clinician, 2:44) — in the export file · approved by Dr Tobin Hale (clinician) · checked against US Core

## Plan

- Draft order: acetaminophen — 500 milligrams by mouth every eight hours as needed (clinician, 3:39) — in the export file · approved by Dr Tobin Hale (clinician) · checked against US Core
The whole note
# Visit note

Export file: 7 approved record(s), with the patient, clinician and visit.

## Subjective

_Symptoms and denials stay in this note; they are not exported._

- Symptom: “I have had a cough for four days. It keeps waking me up, and I am supposed to be back at work tomorrow. I thought I had better get it checked” (patient, 0:13) — note only · approved by Dr Tobin Hale (clinician)
- Symptom: “No fever. I checked a couple of times at home because I wondered about that too” (patient, 0:40) — note only · approved by Dr Tobin Hale (clinician)
- Allergy: sulfa (unconfirmed); stated reaction: rash (patient, 1:57) — in the export file · approved by Dr Tobin Hale (clinician) · drug family, kept as words · checked against US Core
- Allergy denial: “I am not allergic to penicillin. I have taken that before without any trouble” (patient, 2:18) — note only · approved by Dr Tobin Hale (clinician)
  - Note: The installed code set cannot represent this allergy denial as a record

## Objective

- Blood pressure 126/78 mmHg (clinician, 2:36) — in the export file · approved by Dr Tobin Hale (clinician) · checked against US Core
- Heart rate 84 /min (clinician, 2:44) — in the export file · approved by Dr Tobin Hale (clinician) · checked against US Core

## Assessment

- Patient-reported: asthma (unconfirmed, active) (patient, 1:01) — in the export file · approved by Dr Tobin Hale (clinician) · kept as words, no code · checked against US Core
- Family history (not the patient's condition): “My father has diabetes. He has been telling me I should find a regular doctor instead of just coming in when I am sick” (patient, 1:11) — note only · approved by Dr Tobin Hale (clinician)
- Assessed: acute bronchitis (provisional, active) (clinician, 2:58) — in the export file · approved by Dr Tobin Hale (clinician) · kept as words, no code · checked against US Core
- Ruled out: pneumonia (refuted) (clinician, 3:11) — in the export file · approved by Dr Tobin Hale (clinician) · kept as words, no code · checked against US Core

## Plan

- Draft order: acetaminophen — 500 milligrams by mouth every eight hours as needed (clinician, 3:39) — in the export file · approved by Dr Tobin Hale (clinician) · checked against US Core

## Also discussed

_Kept in this note, not exported as records._

- Continuing medicine (not a new order): “I still use albuterol. I brought the inhaler with me because I can never remember everything on the label” (patient, 1:35) — note only · approved by Dr Tobin Hale (clinician)

It ends with a short footer naming the model and the time it was written.

Quick start

Requirements: macOS or Linux, mise, llama-server from llama.cpp on your PATH, and about 10 GB of free disk space.

git clone https://github.com/supermax01/auris.git
cd auris
mise install                              # Python 3.12, uv and Java 21
uv sync
uv run auris setup fhir                   # FHIR packages
uv run auris setup terminology --fetch    # RxNorm and ICD-10-CM
uv run auris setup validator --fetch      # the official HL7 validator
uv run auris setup model --fetch          # the local language model (about 5 GB)
uv run auris demo                         # a fictional visit, written to ./out/demo

Setup is the only step that downloads anything. Every download is checked against a pinned checksum.

Usage

Desktop app

The desktop app needs Node.js 22 or later and Rust. To try it with a fictional clinic day:

npm --prefix ui ci
npm --prefix desktop ci
cd desktop && AURIS_DEMO=1 npm run dev

Pick a visit, choose Start the visit, then Use the fictional demo transcript.

Without Rust, you can run the same interface in your browser:

npm --prefix ui ci && npm --prefix ui run build
uv run auris app serve --demo --data ~/auris-demo

Then open the http://127.0.0.1:… address it prints. Keep that address private: it contains the session key.

Command line

Write a note from a transcript and visit details (see examples/ for the file formats):

uv run auris run --transcript examples/transcript.json \
  --intake examples/intake.json --out out/visit

Review the result in your browser:

uv run auris review serve --run out/visit

Turn a recording into a transcript (one room microphone, or one file per speaker):

uv sync --extra speech
uv run auris setup asr --fetch
uv run auris transcribe --audio visit.wav --out out/visit/transcript.json

Each run writes note.md (the visit note), bundle.json (the FHIR records), extraction.json and validation.json into its own private folder.

Command Description
auris setup <part> Download and verify a component: fhir, terminology, validator, model, asr
auris demo Run the fictional demo visit
auris run Write the note and records for one visit
auris transcribe Turn audio into a transcript
auris review serve Review a visit in the browser
auris app serve Start the engine behind the desktop app
auris fhir-check Load an export into a local FHIR test server and check it

How it works

recording ─► transcript ─► facts ─► checks ─► FHIR records ─► clinician review ─► export
            (Whisper)     (local    (quotes,   (US Core,       (approve, correct,   (FHIR file
                           model)    numbers,   validated       or dismiss)          + note)
                                     speakers)  offline)
  1. Transcript. Whisper turns the audio into text with timestamps, and each speaker is identified as clinician or patient.
  2. Facts. A local model reads the transcript and fills in a fixed form for each fact, quoting the words it relied on.
  3. Checks. Each quote is matched against the transcript. Facts with unsupported numbers, the wrong speaker or a missing negation are flagged or refused.
  4. Records. Facts are converted to FHIR resources, coded from the terminology files, and validated with the official HL7 validator.
  5. Review. The clinician approves, corrects or dismisses each fact. Only approved records can be exported.

Project status

Auris is pre-alpha. Today it can:

  • write a note and FHIR records from a transcript or a recording
  • show every fact with its source for review, in the desktop app or a browser
  • export approved records and the note

Roadmap

Planned work, not built yet:

End-to-end SMART on FHIR integration

Today Auris exports a FHIR file that you load into your EHR yourself. The next major step is to connect it to the EHR directly with SMART on FHIR, the standard way apps plug into Epic, Oracle Health (Cerner) and other FHIR-enabled systems:

  • Launch from the EHR. Open Auris from inside the patient's chart (SMART EHR launch) or on its own (standalone launch), with sign-in through the EHR's OAuth 2.0 and OpenID Connect.
  • Visit details from the chart. Take the patient, the encounter and the clinician from the launch context instead of typing them in.
  • Write back what was approved. After review, send the approved records (conditions, vital signs, allergies, draft medication orders) to the EHR through its FHIR API, using only the permissions the clinician grants. Orders stay drafts for the clinician to sign in the EHR.
  • Still local first. Audio and transcripts stay on the clinician's computer. Only the records the clinician approved are sent, and only when they choose to.
  • Tested end to end against public SMART sandboxes before any real EHR.

Also planned

  • Live listening in the desktop app (today it works from transcript files)
  • A packaged, signed desktop app (today it runs from source)
  • Sign-in and support for several clinicians on one installation

Repository layout

Path Contents
src/auris/ The Python engine: speech, extraction, checks, FHIR mapping, review
ui/ The app interface (React)
desktop/ The desktop window (Tauri)
config/ Pinned versions and checksums of every model, package and terminology file
examples/ A fictional transcript and visit details
docs/ FHIR notes, the US Core profile checklist and the brand assets
tests/ The test suite (all data fictional)

Development

uv run pytest                  # tests that need models or the validator skip until setup
uv run ruff check src tests
uv run ruff format --check src tests

Issues and pull requests are welcome. Please keep all test data fictional.

Privacy

Auris has no telemetry and makes no outside network connections after setup; its parts talk to each other only on your own computer. Visit data is stored in a private folder on your computer, readable only by your user account.

License

Auris is released under the Apache License 2.0.

It builds on Whisper, Qwen3, llama.cpp, SpeechBrain, HL7 FHIR and Tauri. RxNorm, ICD-10-CM, LOINC and UCUM content is used under the terms listed in NOTICE. Terminology files, FHIR packages and model weights are downloaded from their publishers at setup and are not included in this repository.

About

Open-source clinical scribe that runs on your own computer. Turns a doctor–patient visit into a quote-backed note and FHIR records; the clinician approves every fact.

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages