Is Google’s NotebookLM Going to Disrupt the Podcasting Industry? | by Dr. Varshita Sher | Oct, 2024

Generative AI

For those who may not know, NotebookLM is a personalized AI research assistant powered by Gemini 1.5 Pro, designed to make sense of complex information. In addition to answering questions based on your uploaded sources (documents, slides, charts, etc.), it can also create personalized study materials by automatically generating things like a table of contents, study guides, briefing documents, FAQs, and more. While it formulates answers based on the uploaded sources, it also provides inline citations, highlighting the specific text blocks in the source documents used to generate the response.

The uploaded content can range from research papers and meeting transcripts to quotes from interesting books, chapters of a novel you’re writing, corporate documents, and more. These sources can include Google Docs, Slides, PDFs, text files, copied text, and even web pages.

Now, to the main reason for this article: Last month, NotebookLM announced a new feature — Audio Overviews — which has been making headlines. This feature offers a new way to interact with your source documents. With just one click, it generates engaging “deep dive” discussions that summarize the key topics in your sources.

What’s even more impressive is how it transforms any piece of content, no matter how dry, by generating two AI hosts (one male and one female) who discuss the document’s contents in a podcast-style format.

If you’re wondering what “podcast-style format” means, imagine the friendly banter, the little jokes, the back-and-forth conversations, the laughs, interruptions, “umms,” and “you knows’”— essentially all the hallmarks of a great podcast listening experience.

These podcast-style conversations create natural connections and segues from your text, resulting in an engaging dialogue.

To test it out, I decided to repurpose one of my old Medium articles and create a podcast from it to cater to a more audio-loving audience.

The set up for the same was quite straightforward.

• Go to NotebookLM. You’ll have to sign in with your Google ID if you aren’t already. If it’s your first visit, you’ll see several sample notebooks and you can create a new one with the “Create” button.
• Next, add content to your notebook. I used the website source to feed in my Medium article. Alternatively, you can paste text or fetch from Google Drive.

• Finally, click the “Generate” button inside the Notebook guide (see image below) to create the audio. And go grab a ☕️ as it might take a few minutes depending on the content length.

P.S. It took around 4 minutes to generate a 13 minute audio from my 1100-word article. You can play and listen here.

P.S. I ended up trying Audio Overview with various sources, such as podcast transcripts, research papers, and data science blogs. The following takeaways are an amalgamation of my experiences across all these sources.

Let’s start with the good stuff:

• It’s remarkable that we can quickly create a podcast episode in just minutes, allowing many of us to have a side gig as podcasters (should you choose to). This is a great way for writers to repurpose their content and for others to engage with relatively complex topics in a fun and accessible manner.
• The use of analogies throughout the audio is truly remarkable and captivating. In the case of my Medium article, it was able to take a relatively niche (read:boring) topic (scaling challenges with Gen AI might not appeal to everyone outside the immediate field) and make connections to everyday things.
  For instance, at one point the hosts discuss Gen AI token costs and provide a much more relatable example, comparing how these costs can add up to micro-transactions in a mobile game. Similarly, they explain prompt engineering with an example of providing a complete recipe with measurements, rather than simply saying “make me a delicious meal”. They also use the analogy of a car remembering a common route to explain LLM caching.
• The way the two hosts build on each other’s sentences feels very natural, and the segues flow seamlessly. For example, using phrases like “speaking of…” to introduce a new topic feels organic and not forced at all.
• Emphasis on certain words at just the right moments helps hold the listeners’ attention. Expressions like “oh wow”, “oops”, and “aah” convey genuine surprise at what the other host just said. Natural pauses to think of the right word make the conversation feel spontaneous rather than rehearsed.
• After testing this on several deep learning papers, I can confidently say it will be a game changer for explaining complex research that benefits from analogies and “explain like I’m five” (ELI5) examples. In fact, the guidelines in one of their pre-prepared example notebooks, titled Introduction to NotebookLM, state that it’s designed for researchers, journalists, students, and business professionals.

Having looked at the key advantages, there are also a few limitations to consider:

• Sometimes, the conversation between the two hosts doesn’t feel real. Very often, they finish each other’s sentences, even when the first host has just asked the second host to explain a new concept and a few seconds later, Host 1 ends up answering their own question.
• Not all input sources generate audio of equal quality. As part of stress testing, I tried inputting the transcript from another podcast, and the hosts seemed more inclined to make humorous noises at each other — ‘yayaya,’ ‘oh yeah,’ ‘hmm,’ ‘uh-huh,’ ‘right,’ ‘gotcha,’ etc.!
• The only downside to having a lot of analogies while discussing a topic is that sometimes the AI can get the analogies wrong. For instance, while discussing a blog on forecasting metrics, it used the analogy of “just like in schools a lower score is generally better, it means your forecast is closer to reality”.
  Such hallucinations are common across different generative AI models and have been included as a disclaimer in their tool as well. These might be more pronounced if we provide a very niche, highly specialized topic, such as the role of microRNAs in gene regulation (the topic that won the Nobel Prize in 2024 this week). In such cases, it may start hallucinating with analogies used due to a lack of relevant inherent knowledge🤷‍♀.
• For very large texts, the podcast can often end abruptly. This suggests that there may be a cutoff point for the training data, beyond which the audio cannot adapt to provide a smooth, natural ending.
• (Very minor but) Some of the words, mostly abbreviations, are garbled in the audio. For some reason RAG is pronounced as ArrrR-G instead of individual alphabets like R-A-G.
• At times, hosts overly agree with one another, using filler words like ‘right’ and ‘exactly’ while the other host is still talking. This can feel like forced responses; I mean, let the poor guy finish!

Now that we’ve covered the good and the bad, let’s move on to the million-dollar question: is this new tech enough to give podcasters a serious competition?

My simple answer is — not yet. The reason? All the aforementioned issues we’ve discussed. And I know some of you might disagree and say these problems are minor, and you’d be right. If you listen to just one podcast, you may not even notice them, but if you continuously listen to multiple episodes, especially on a daily or weekly basis, the sheer number of analogies and “exactlys” can become overwhelming. For these reasons, perhaps Google never positioned it as a podcasting tool in their initial release.

That said, it will definitely lower the barrier to entry for many who want to explore this field but may not want to use their own voice for various reasons. More importantly, I see its use as a way to consume complex topics in digestible formats.