Skip to content

Glossary

Ambient scribe. Software that records a clinical visit and drafts the documentation from the conversation, as opposed to dictation, where the clinician speaks the note.

ASR / STT. Automatic speech recognition, speech to text. Whisper and Vosk are ASR engines.

BAA. Business associate agreement. In the US, the contract required before a vendor may handle PHI on a covered entity's behalf. The scribe's locality guard exists because of this. Cloud models and PHI lists who offers one.

Beam size. How many candidate transcriptions Whisper keeps while decoding. More is more accurate and slower.

CTranslate2. The inference engine faster-whisper runs on. Handles the GPU and quantisation.

Diarisation. Working out who spoke when. Whisper does not do it; pyannote does.

Hallucination. The model producing fluent text with no basis in the input. In transcription: words in silence. In drafting: findings not in the transcript.

LLM. Large language model. The part that reads the transcript and writes the note.

Local-first. Designed so everything runs on the machine in front of you and nothing requires a network.

OpenAI-compatible API. The HTTP request format OpenAI introduced and nearly every model server now accepts. Why one client talks to Ollama, vLLM, llama.cpp and LM Studio.

PHI / PII. Protected health information; personally identifiable information. Audio, transcript and note are all PHI.

Quantisation. Storing model weights in fewer bits (4-bit, 8-bit) to use less memory and run faster at a small cost in quality.

Reasoning / thinking model. A model trained to write out a hidden chain of thought before answering. Good for puzzles, slow and unnecessary for drafting from a transcript.

Retention. How long each kind of data is kept before deletion. Decided before the first visit, enforced by the review app each time it starts and hourly while it runs, or by scribe purge on a schedule.

Segment. One chunk of a transcript with a start time, end time and text.

Session. One visit's folder: audio, transcript, note, metadata.

SOAP. Subjective, Objective, Assessment, Plan. The most common clinical note structure.

System prompt / template. The instructions sent to the language model before the transcript. In this project, a Markdown file.

Temperature. The randomness setting on a language model. Low is repeatable and literal; high is varied. A scribe wants low.

VAD. Voice activity detection. Finding the parts of audio that contain speech so the rest can be skipped.

VRAM. Memory on the graphics card. The limit on which models fit.

Whisper. OpenAI's open-weights speech recognition model family. The default ASR in this course.