ReadImaging

Lightning-Fast & HIPAA Compliant Radiology Dictation.

Impressed? Let's talk. [email protected]

Sales demonstration only. Not for clinical use.

Initializing...

Downloading speech recognition model...

Sample Scripts

Select a script to read along while dictating

Records Rocket SendImaging ReadImaging

How ReadImaging Dictation Works

This dictation engine is not a large language model (LLM). It does not generate, predict, or hallucinate text. It maps sound directly to words.

No hallucination risk. Unlike LLM-based systems (including OpenAI Whisper), this model cannot invent words, fabricate findings, or insert content that was never spoken. Every word in the transcript came from your voice.

The Architecture: Conformer CTC

The model uses a Conformer architecture — a hybrid that combines Transformers (which understand long-range context across a sentence) with Convolutional Neural Networks (which extract precise, local audio features like individual syllables).

These two technologies are layered in a repeating "sandwich" structure that gives the model both the big-picture understanding and the fine-grained acoustic precision needed for medical terminology.

Why It Can't Hallucinate

The model uses CTC (Connectionist Temporal Classification) decoding. This means it classifies each tiny slice of audio independently — it does not auto-regressively generate the next token based on previous output.

In contrast, Whisper and other transformer-based ASR models generate text one word at a time, where each word depends on the words before it. This auto-regressive loop is what causes hallucination — the model can "get creative" and generate plausible-sounding text that was never spoken.

CTC models cannot do this. They are constrained to only produce output that directly corresponds to audio input.

Trained on Medical Speech

The model was pre-trained on approximately 5,000 hours of de-identified physician dictations across multiple specialties, including radiology, internal medicine, and family medicine. It recognizes complex anatomical and pathological terms out of the box.

100% Local Processing

The entire model runs in your browser via WebAssembly. Audio is processed in a zero-copy ring buffer — it never touches a server, is never uploaded, and is never stored. When you close this tab, the audio is gone.