NarrateRad — Dictation That Catches Your Mistakes

Watch

The 36-second tour

Demo dictation of a fictional post-procedure CT. The flags and the report are the app's real output.

The Problem

Fast dictation, expensive errors

Dictation is the fastest way to report, and the easiest place for small errors to hide: the wrong side, a finding denied in one sentence and described in the next, a critical result buried in line nine. Conventional speech recognition transcribes faithfully — including the mistakes — and leaves catching them to the person who just made them.

What I Built

A second pair of eyes that listens while you dictate

NarrateRad streams audio from the browser, transcribes it with Whisper, and runs rule-based clinical checks on every chunk as it arrives. Low-confidence words are highlighted for a second look, colloquial phrases are mapped to RadLex terms, and when the dictation is done, Claude organises it into a preliminary report — and points out what still needs confirming.

Technical Highlights

Safety checks that run in real time

Laterality conflicts

If the same structure appears with both “left” and “right” — and isn't described as bilateral — the dictation is flagged before it becomes a report.

Negation contradictions

A finding that is both negated and affirmed (“no pneumothorax” … “small right pneumothorax”) raises a contradiction flag.

Critical findings banner

Pneumothorax, haemorrhage, cord compression, free air and similar results trigger a red banner prompting confirmation and escalation.

RadLex standardisation

Plain phrases like “big liver” or “fluid around the lung” are mapped to standard terms — hepatomegaly, pleural effusion.

No patient identifiers

Name, MR number and date-of-birth fields were removed from the app entirely, with a reminder not to dictate identifiers.

Two ways to run

A local desktop app with on-device Whisper on Apple Silicon, and a zero-backend web build where each user brings their own API keys — so hosting costs nothing at any traffic level.

Stack

Under the hood

Transcription	Whisper large-v3 — mlx-whisper on-device on Apple Silicon, Groq API elsewhere
Server	Python + FastAPI, WebSocket audio streaming
Clinical NLP	Pure rule-based checks — laterality, negation, opposing descriptors, critical findings
Reports	Claude, for radiology and histopathology
Web build	Static HTML/JS, no backend; installers for Mac and Windows

Status

Where it stands

Working and installable with one-line installers for Mac and Windows, with a self-update mechanism. The static web build is ready to host. Transcription stays on-device only on Apple Silicon; report structuring always uses the Claude API.