How to summarise medical lectures with AI: a four-step pipeline that turns one hour of audio into a study guide, flashcards, and an audio overview

Why a structured pipeline beats 'summarise this for me'

Throwing a lecture transcript into ChatGPT and asking for a summary is the workflow most students try first. It is also why most students stop using AI on lectures after a fortnight. The output reads well but contains invented drug doses, fabricated trial names and confidently stated guideline numbers that do not exist. For exam revision, that is worse than no summary.

A structured pipeline fixes this by separating the jobs: a transcription model that does only speech-to-text, a retrieval-grounded model that summarises only from the transcript, and a generation model that drafts cards from the verified summary. Each step has one job, and the citations stay attached to the source. The 15 minutes of work returns assets you will actually revise from.

Step 1: record the lecture (and stay on the right side of policy)

Always check your institution's recording policy first. Many UK and US medical schools record lectures centrally and let you download from Panopto, Echo360, Zoom Cloud or the VLE; use that recording where you can because consent is already covered. If you are recording yourself, ask the lecturer and disclose that the audio will go through an AI tool. Audio-only is fine for this pipeline; video adds size without value.

For clinical teaching at the bedside or in theatre, do not record. Patient identifiable data is in the room, and most trusts in the UK and HIPAA-covered entities in the US treat that as a hard line. Take written notes for those sessions and apply the rest of the pipeline only to the formal didactic content.

Step 2: transcribe with Whisper

Whisper is OpenAI's open-source speech recognition model, released in 2022 and updated through the large-v3 release. It handles medical terminology better than almost any other free transcription tool because its training data included a wide swathe of professional speech. Most modern phone voice memo apps now use a Whisper-class model under the hood, so on iPhone or Pixel the transcription is already built in.

On a laptop you can run Whisper locally for free via tools like MacWhisper, Whisper.cpp or the openai-whisper Python package. Cloud wrappers like Otter.ai and Notta layer extra features on top. A 60-minute lecture transcribes in roughly three to six minutes on an M-series Mac and produces a plain text file with rough timestamps. Do not bother cleaning the transcript; the next step does not care about formatting.

Step 3: distil the transcript in NotebookLM

NotebookLM is Google's source-grounded research assistant. It only answers from documents you upload, and every answer cites the exact paragraph or timestamp. Upload the Whisper transcript and, if you have them, the lecturer's slide PDF and any handouts. Pinning multiple sources to one notebook lets the model triangulate between what the lecturer said and what was on the slide.

Run the prompts in this order, in the same notebook. Each one builds on the last.

1. 'Produce a study guide for this lecture organised by topic. Include the three highest-yield facts per topic and one common exam trap.'
2. 'List the ten most likely exam questions from this lecture and answer each in one to two sentences, citing the timestamp.'
3. 'Compare what the transcript says with the slides. Flag any contradictions.'
4. 'Generate a 10-minute audio overview suitable for revision on a commute.'
5. 'List every drug, dose and guideline mentioned, with the exact slide or timestamp.'

The audio overview is the underrated feature. NotebookLM's Audio Overviews produce a two-host conversational podcast from your sources, typically eight to twelve minutes long, that catches connections you would skim past in text. For deeper coverage of the tool itself, see notebooklm-for-medical-school.

Step 4: convert the study guide into Anki flashcards

Take the study guide and the exam questions from step three, paste them into ChatGPT and ask for atomic cloze deletion cards in Anki import format. The full prompt is in chatgpt-for-medical-students. Aim for 30 to 50 cards from a one-hour lecture; more than that is usually noise rather than signal. Import the deck into Anki, tag it with the module and date, and you are done.

If the lecture covered anatomy, surgery or pathology with images, generate the diagrams separately with a sketch-first illustration tool such as Angiosome and use Anki's image-occlusion add-on. Pure ChatGPT image generation is not reliable for anatomy because it hallucinates structural detail. See how-to-illustrate-medical-notes-with-ai for the diagram workflow.

A worked example: a 55-minute heart failure lecture

Cardiology, second-year clinical, 'Heart Failure: Classification and Management'. The lecture was on Panopto so consent was already covered. Total active AI time: 22 minutes. Total passive AI time (audio generation, transcription run): 18 minutes that overlapped with making dinner.

1. Downloaded the Panopto MP4. Transcribed with MacWhisper large-v3 in 4 minutes 12 seconds.
2. Uploaded transcript and slide PDF to a new NotebookLM notebook called 'Cardiology > Heart Failure'.
3. Ran the five prompts. Got a 1.5-page study guide, 14 likely exam questions with answers, three flagged contradictions between speaker and slides, and an 11-minute audio overview titled by the model 'NICE 2024 update: what changed'.
4. Pasted the study guide into ChatGPT with the cloze prompt. Got 38 cards. Spot-checked the five drug-dose cards against the BNF.
5. Imported into Anki under 'cardiology::heart-failure::2026-lecture-04'. Listened to the audio overview at the gym that evening.

Pitfalls and how to dodge them

• Transcript errors on drug names. Whisper occasionally mishears 'losartan' as 'lozartan'. Manually verify pharmacology cards against the BNF.
• Lecturer asides treated as content. NotebookLM does not distinguish jokes from facts. Skim the study guide once and delete the noise.
• Image-only slides. The model cannot read those. Either type a one-line description into the speaker notes before upload, or annotate them in your own notes.
• Patient cases mentioned in lecture. Even anonymised cases can be re-identifiable. Redact specific dates, units and locations before upload, or skip those sections.
• Volume cosplay. 200 cards from one lecture is not better revision; it is a deck you will abandon. Cap at 50 and prefer quality.
• Over-trusting audio overviews. The two-host format is engaging, which makes errors land harder. Treat as passive reinforcement, never as primary learning.

Safety, consent and academic integrity

Three rules cover almost every problem. Do not record without consent. Do not upload identifiable patient information to a cloud AI tool, including dates and locations that could re-identify a case. Disclose AI-generated study material in any formal submission if your school's policy requires it; many UK schools follow the GMC's 2024 generative AI guidance and most US schools follow AAMC and institutional honour codes.

Personal revision notes are normally outside the disclosure requirement; portfolios, reflective writing and assessed coursework are not. When in doubt, declare. Keep the original transcripts and the NotebookLM notebook URL as your evidence chain.