NotebookLM is a great tool for generating ideas, summarizing content, and providing various ways to study certain material. But when I learned about its Audio Overviews feature, which creates AI-generated podcasts from your sources, my expectations were low.
However, once I used the feature, I realized why it's one of the reasons to use NotebookLM. While the format may not suit every use case, it's a great way to quickly dive into a topic that relates to your sources.
Related
I finally started using NotebookLM and I should have sooner
I'm officially a NotebookLM convert
Posts 4
Why I was skeptical of Audio Overviews
I'm not a big fan of generative AI
While text-to-speech technology has come a long way since its early iterations, the quality varies greatly depending on the source of the speech synthesis. Text-to-speech in your computer's accessibility settings, for example, may sound very robotic (as was the case when I tested speech properties on my Windows 11 PC).
Even for AI that is supposed to sound more natural, such as the AI-generated voice-overs that can be found in some YouTube videos, the voice often sounds strange. The pacing, awkward intonation, and mispronunciation of certain words can often give away the AI-generated nature of the voice-over.
This isn't limited to speech but also to text produced by generative AI. Many of us are able to spot AI-generated writing on social media and websites, even if the actual language used is grammatically correct.
I had heard great things about NotebookLM but mostly dismissed the praise as over-hyped. After all, my experience with AI tools mostly consists of high hopes and eventual disappointment. However, audio overviews are one of those features that NotebookLM does better than competitors.
How Audio Overviews exceeded expectations
Artificial speech that sounds surprisingly natural
Close
I tested Audio Overviews on a variety of topics and found that the results far exceeded my expectations. The hosts sound natural, seemingly react to what the other says, and mimic human speech by including pauses, filler words like "uh," and even the occasional stutter.
While you will always get two hosts, one male and one female, they adapt to the sources you use and the tone of the topic. For example, for an overview of a medical topic, the hosts sound more serious and sympathetic. The hosts sound more upbeat for a more casual topic, like tech trends or how to tame a certain dinosaur in the game Ark: Survival Evolved.
NotebookLM generates its output based on the sources you provide it. This helps you avoid inaccurate information by limiting the sources to the sites and content that you trust. While I was able to spot one error in NotebookLM's Timeline feature, its text output and Audio Overviews feature have never generated AI hallucinations for me -- another frequent drawback I encounter when using AI tools.
If you're trying to approach a topic in an approachable way, I'd recommend trying out Audio Overviews. The feature is included in NotebookLM's free plan, though you can only generate a certain number of audio recordings in a single day.
Even non-English overviews proved impressive
A rare feat for a lot of AI
Audio Overviews also have a few features still in beta -- namely Interactive mode and generating non-English podcasts. Since accurate language support is something I rarely see in AI tools, I wanted to try out the feature in another language.
While there's no way to change a specific Audio Overview to another language, you can change your default output language for NotebookLM. This allows you to generate new podcasts in the selected language then.
To change your output language, select Settings -> Output language and select from one of the supported languages. While generative AI tools often only support a few languages, NotebookLM supports Audio Overviews for dozens of languages, including Arabic, Chinese, Danish, Filipino, Spanish, Swahili, and more.
I selected Afrikaans for my output, as this is my second language. I then re-generated my Audio Overview on fibromyalgia to see how accurate the translation was, as well as how natural the accents of the hosts would sound.
While the Afrikaans accents of the hosts felt a bit less natural than their English counterparts, they still sounded proficient in the language. I also couldn't spot any errors in the actual vocabulary or grammar of the hosts.
To double-check their fluency, I also asked a friend to listen to the generated podcast to see if he could spot any errors, as Afrikaans is his home language. He mostly spotted a few points where the pronunciation was off, as well as a few instances of direct translations that didn't sound quite right. Overall, though, the recording was good -- just not perfect. But I was surprised at how well the hosts performed for a language spoken by roughly 7.5 million people.
Related
4 productivity tricks I've learned after using NotebookLM since launch day
These simple yet powerful NotebookLM tricks have transformed my workflow.
Posts 7
I'm interested to see where Audio Overviews go next
Audio Overviews are one of the most promising features of NotebookLM, but it also depends on how people use it. I don't think it would serve as a real (or ethical) replacement for podcasts by actual people. But it is certainly a great way to dive into topics in a quick and engaging way. I would like to see transcripts added to the feature; however, I think it's one of the features NotebookLM should introduce.