How to Make a Medical Podcast Using AI (Part 2)

How to Make a Generative AI Medical Podcast (That’s Actually Accurate)

The original hope was that using AI might allow us to create a single podcast episode in one afternoon. That’s much faster than the time it would take for someone to do everything on their own: research and write scripts for an episode, record it, edit it, and so on.

But things didn’t go quite the way we planned.

So, here’s what we’ve tried so far, what worked and what didn’t, plus, what we’re hoping to accomplish in future.

Finding the right AI to “host” the podcast

Wondercraft AI is the service we settled on because they have high-quality voices that aren’t robotic; plus, the platform can do everything including generating scripts, adding music, and creating audio in the correct format for publication. But we didn’t end up using all of those features because we got better results for certain parts of the process by bringing in other tools, which we’ll discuss below.

The first task was to choose AI voices to represent the podcast hosts, the fictional “Nora” and “Peter. Next, we had to generate the scripts.

Using generative AI to create error-free scripts for a medical podcast

Franz selects a few studies to discuss in each episode. We feed the most important takeaways into the AI, and prompt it to create a podcast script for us that’s based on the info we provided. 

We’ve found it easier to work with ChatGPT4 to generate the scripts, rather than using Wondercraft’s built-in script generation. This is largely because ChatGPT allows us more flexibility with the types of instructions (prompts) that we give it.

But the script isn’t finished yet…

Next, we go through the script to fine-tune it and make it sound the way we want. At the moment, we’re admittedly still leaving some of the AI’s default “personality” in the script (i.e. we’re not necessarily editing it to make it sound exactly like what we’d have written on our own).

Next, we fact check the script line-by-line, to ensure the AI’s output is in agreement with the info we provided to it in the first place. If any errors are present, we correct them.

(By the way, if you want help creating effective prompts, grab a trial account at Medmastery and use our AI prompting tool. You’ll also get access to our entire ChatGPT course!)

Generating audio for the podcast

Now that we have our script, we head back to Wondercraft to generate the audio. However, we ran into a challenge…

Pronunciation problems

The AI commonly mispronounces medical terms. And if it’s narrating a serious topic, it’ll often sound overly dramatic or excessively solemn. In such cases, we regenerate the problematic audio until it’s as close to perfect as possible. 

One tactic we use to make the AI pronounce things correctly is to adjust the spelling within the script, or split one word into two. For example, in a recent episode discussing dapagliflozin and anemia in patients with chronic kidney disease, the AI often pronounced “dapagliflozin” very strangely. To fix it, we had to change the spelling of “dapagliflozin” in the script to “dapa glifloe-zin”.

Another challenge

It’s common for the AI to add new errors in pronunciation when we regenerate an audio segment to fix a different pronunciation issue. 

Here’s the best workaround we’ve found so far:

1. Export the original audio segment from Wondercraft (with errors). 
2. Then return to Wondercraft to fix the mispronunciations. We experiment with different spellings for the problematic words, and regenerate the audio either for the entire sentence or just a phrase that contains the problem word, until the AI voice finally says it correctly. 
3. Once we succeed, we export the corrected piece of audio.
4. We use software capable of editing audio to splice in the corrected words or sentences.

We’ve found that method to be faster than trying to get the AI to correctly pronounce an entire minutes-long segment in one go. 

Audio editing software

We’re using Camtasia because that’s what we had handy and felt most comfortable using as an audio editor. Then we use Audacity to quickly convert the output to an .mp3 file format. 

If you’re already proficient in editing audio with Audacity (or some other program), there’s no advantage to using Camtasia at all—in fact, you can use any audio-editing program you like. Whichever audio editor you use, the key is that the final audio file should be an .mp3 file, as that’s the most widely supported format.

Distributing the podcast

The last step is uploading the audio to a website called Libsyn. It’s a platform that distributes the audio to popular podcast streaming platforms such as Spotify and Apple Podcasts. This is also where we input the podcast and episode descriptions that appear on streaming platforms.

How much does creating a generative AI medical podcast cost?

Here’s a summary of our costs for the Cardiology Digest podcast:

If you wish, you can do it even more inexpensively than we did. For example, instead of using Libsyn for podcast distribution, you may find that the free tier of something like Podbean works just fine for your needs. You could use a free software like Audacity for all of your audio editing. And you could use the free version of ChatGPT to help you with scripts.

Want to check out our AI podcast and see what you think? You can listen to all our episodes here. 

References

Educational Resources

• Medmastery Cardiology podcasts [Spotify] [Apple podcasts]
• Guilleminot S. Making a medical podcast using AI – Part 1. LITFL
• ChatGPT Essentials for Clinicians. Medmastery
• Guilleminot S. AI in Healthcare. LITFL