Write a presentation/conference talk by speaking

Summary

Writing a presentation or conf talk takes hours - but can we 'speak' it and have the AI invent the slides, create the structure, write the speaker notes?

Two challenges in conference talks

The reality of writing a conference talk is that it takes a professional about 8x as long to prepare the talk as it takes to deliver it – i.e. a 1 hour talk requires approximately 1 full day of work to create. Day-to-day presentations (e.g. all-hands meetings, e.g. new sales pitches) have a similar ratio. So the ‘speaking’ part of speaking is quick but the ‘behind the scenes’ is painfully slow / time-consuming.

Separately: when we reason orally (speaking out loud, to other people) our brains construct different liens of argument to when we reason in writing (either digital typing or handwritten on paper). Since a talk is delivered orally there is a significant disconnect if you try to construct it without speaking – and the quality suffers / requires more practicing to test and fix all the problems caused by starting it by thinking ‘on paper’.

Traditional Approach

Classic ways to write a new presentation

1. EITHER: Create a slide deck, write the ‘key points’, gradually add more and more slides until they’re fully covered. Then try presenting the deck to an empty room – typically 5 or more times – recording it each time, and listening to each recording to identify the glaringly bad parts so you can re-think / fix them.
2. OR: Present to an empty room – typically 3 or more times – stopping to write notes every time you ad-lib a particularly good turn-of-phrase or good point. Take your notes and try to wrangle them into a slide-deck format. Then re-present – typically 2 or more times – to correct the bad parts that have crept-in.

Improvement: Use AI and direct-to-Audio to remove multiple steps

Since 2023 we’ve had excellent audio capabilities in the modern LLM-based AI (ChatGPT, Gemini, etc). This is used heavily for interactive conversations over the phone with an LLM (since early 2024) – but also for high-quality low-cost transcription (audio converted to text).

IDEA: Can we speak the presentation once, have an AI listen to it – and then make the AI do everything else for us?

Provisional steps:

1. Deliver the presentation informally, speaking to an AI
2. Use an AI (same or different) to examine the presentation and craft a narrative structure
3. Use an AI to create a slide deck (titles, bullets) from the narrative
4. Use an AI to surface key points and phrases to insert as Speaker Notes

Concrete implementation

Step 1: Speak to an LLM

All the LLMs now have cheap-or-free audio transcription – but in most cases the vendors make it artificially difficult to access if you want direct access. It’s only 5 lines of code, but for some reason they make you write it instead of have it as a simple built-in option. I couldn’t be bothered to open a code editor for this, so instead I used https://whispering.bradenwong.com/ which lets you speak to ChatGPT’s backend and it outputs the text when you’re done.

NOTE: Braden’s site requires you to go to your ChatGPT API page and generate a new API key, and copy/paste it into the site. This is sensible, it means it costs Braden almost nothing to run the site, because it’s directly using your OpenAI account/credits. I highly recommend generating a single key for this, and when you’re finished if you have any fear about Braden misusing your account acccess … simply delete the key. You’re guaranteed no-one has access to your account after that. Much simpler and safer than credit cards etc.

Steps 2 and 3: Prompt an LLM to write slides

I used this prompt with ChatGPT-4o (didn’t bother with the newer models), and then copy/pasted the transcript from Step 1:

Read this text and rewrite it into a conference-talk. Detect and preserve the narrative structure – this is essential – but figure-out the salient slide titles and headings. The style of talk we want is: punchy, rapid-fire, inspirational — but also: contains clear, identifiable, memorable practical tips that stand out to the audience and they remember them afterwards.

Finally: estimate durations for each slide, and a total duration for the talk.

Output looked like:

Step 4: Final deck

I then used https://www.canva.com/ to create the slides, copy/pasting from the clean text that ChatGPT had generated.

NOTE: this step embarrassingly was where things went from “super fast” to “super slow”: commercial vendors of Presentation software are stuck in the last millennium. Several (Canva included) don’t support basic operations like “copy/paste text in as slides” (at least: I couldn’t figure it out via menus + googling). So I had to make the slides individually by copy/pasting one title, one bullet list, at a time. I’ll use something else next time – but they each bring different frustrations and missing/broken core features.

Overall

• OpenAI Whisper via Braden’s website? – excellent
• ChatGPT converting transcript to slides + content? – excellent
• Creating a slide deck from literal “titles + bullets”? – fail; need to find better presentation software

So … could do better … but it wasn’t the AI that let me down, it was my go-to desktop/web software vendors.