Google’s Audio Overview can create podcasts. We put it to the test
Google’s AI note-taking product, NotebookLM, has launched a new feature that allows users to turn their documents into audio discussions between two AI “hosts”.
Audio Overview generates a “deep dive” discussion based on the sources in your notebook, including uploaded files, text or links to websites.
The AI hosts summarise the materials and are capable of making connections between topics. They also banter back and forth in realistic ways.
In a blog post announcing the feature, Google Labs project manager Biao Wang said: “It’s important to remember that these generated discussions are not a comprehensive or objective view of a topic, but simply a reflection of the sources that you’ve uploaded.”
He added that Audio Overview is “still experimental and has some known limitations” — for example, it can take several minutes to generate audio for large notebooks, the AI hosts currently speak only in English and they “sometimes introduce inaccuracies”.
When using NotebookLM, text appears at the bottom of the screen stating: “NotebookLM may still sometimes give inaccurate responses, so you may want to confirm any facts independently”.
Google has already launched a similar non-audio tool for its search function, AI Overviews, which debuted in the US and rolled out to more markets last month.
Analysis: Testing reveals decent accuracy but uncanny valley
The Media Leader decided to test the feature using a variety of text and links added to NotebookLM.
Results were impressive, with fairly strong levels of accuracy and real-sounding conversation between the male and female “hosts”.
There could be various use cases for the tool, apart from actually publishing fully AI-created podcasts. For example, while Audio Overviews do not currently come with a supplied transcript of the audio, there’s a clear potential for it to be used to help script human-recorded podcasts or otherwise develop useful summaries of notes and briefs.
Example 1: Lincoln’s masterpiece
When tested, Audio Overview was able to recognise famous text and provide additional details and context around it, even if it was not fed that additional information.
For example, when The Media Leader fed NotebookLM only the text of Abraham Lincoln’s Gettysburg Address, the Audio Overview “hosts” were able to identify the text and go back and forth about not just the content and brevity of the speech, but additional information about when and why it was delivered amid the height of the US Civil War.
Listen here:
Example 2: Google vs Google
The Media Leader also separately fed NotebookLM a number of articles (from The Media Leader and other sources including The New York Times, The Verge and CheckMyAds) related to Google’s ongoing adtech antitrust trial occurring in the US, leading to an output of a 12-minute episode on the topic (a clip of which is provided below).
The Audio Overview produced an accurate summary of the articles it was fed and, therefore, an accurate summary of the key facts of the trial. Indeed, the result sounded not too dissimilar from The Media Leader Podcast‘s own output earlier this week.
It also included a “jargon-busting” segment where the hosts broke down some of the technical terms related to Google’s adtech stack and a call to action urging listeners to do more of their own research about the trial and its implications.
Listen here:
Example 3: The go-to source (their words)
When given seven links to pages from The Media Leader, the Audio Overview feature produced a six-minute clip of complimentary summaries about the publication.
Indeed, despite not asking for the “deep dive” to feature promotional material, the content is framed as almost an advertisement for The Media Leader. For example, it begins with one of the “hosts” asking: “Ever wish you had like, uh, a cheat-sheet for the UK media scene? Like who’s who, what’s what — all that? Well, that’s what today’s deep dive is all about. We’re taking a good hard look at The Media Leader.“
The female host then replies: “They’re kind of the go-to source for anyone who really wants to understand what’s going on in UK media and advertising.”
Listen here:
There are some inaccuracies. The Media Leader editor-in-chief Omar Oakes is described as a “leading voice” at the Telegraph and Campaign. While the latter is true (Oakes served as media editor and news editor at the Haymarket publication before joining The Media Leader), the former isn’t quite: Oakes worked at the Telegraph for just 10 months as a deputy iPad production editor.
However, the audio does provide accurate summary details about The Media Leader‘s recent content.
Humorously, at the end of the clip, the Overview parrots a standard podcast outro: “That’s a wrap on another deep dive. But hey, we want to hear from you. What insights from The Media Leader really got you thinking? Head over to our website, find us on social media — whatever’s easiest — and let’s keep this conversation going.”
Of course, Audio Overview does not have a website or social media where listeners can “keep this conversation going”.
Example 4: Let’s get meta
Finally, The Media Leader uploaded a draft of this very article into NotebookLM.
It then generated the following summary — so if you’d prefer to listen to what this article is about rather than read it, here it is.
Note how the AI also provided additional information about how large language models work (something not included in this article) and made this judgement on the error regarding Oakes’ career history: “Is a little flub about a career history really the end of the world?”
The summary concluded with pros and cons of the tech, reminding listeners: “We can’t take everything it says as gospel — always double-check, use your judgement, be sceptical.”
Listen here:
À lire plus tard
Vous devez être inscrit pour ajouter cet article à votre liste de lecture
S'inscrire Déjà inscrit ? Connectez-vous