đŸȘŽ Anil's Garden

NotebookLM and Automated Podcasting

06 Jan 20254 min read

NotebookLM Workflow

Upload a single or multiple artefacts. Grab the arXiv HTML link for a paper and provide that directly to NotebookLM via the link option. Where no HTML is available for an arXiv article, you can use ar5iv. NotebookLM “source” upload options include Google Drive, links (scraping handled on their side), or copy/paste of text (e.g. custom docs).

Policies for respect copyrighted material are implemented. I could not upload a Guardian article using the link. Users can and will simply paste in content themselves.

Generated summaries and study guides can be converted to source - a Markdown doc is added to the sources.

Interaction with NotebookLM chunks text and performs RAG with text chunks, providing inline citations (hyperlinking to cited chunks).

Audio overviews: You can share links or download the audio.

Some thoughts

  • The tone is very high-level, which isn’t useful for paper summaries - ideally we’d focus on core contributions. When targeting a summary, isolating the incremental contribution with just enough the background as two distinct things provides a clear snapshot.
  • The repeated platitudes in the “script” diminish quality
  • Interactive mode (beta on trying) was surprisingly great even if the pretence of being a caller into a radio show is weird
    • I asked about the architecture of the brain module (processes EEG/MEG data) and received an in-depth response detailing the 1D convolutional architecture and activation functions (GLU) including the fact that the final convolutional layer does not have a residual connection.
  • Google Illuminate is better for scientists’ use case
  • Google Learn About would be great to try but “Learn About isn’t currently available in your location”

Some hilarious NotebookLM examples are on this Reddit NotebookLM Lounge thread including Genesis for Kids which “[makes] the hosts explain The Book of Genesis for kids, and sing a song about it at the end.”.

Examples

Synthesising News on the UK Horizon Post Office Scandal

Multiple sources integrated from the BBC, Sky News and The Guardian on the recent Post Office Horizon Scandal in the UK.

I used the following articles:

  1. 1,000 days of tears and buck-passing: the evidence from the Post Office inquiry’s key witnesses - Pasted in manually in testing (NotebookLM detects Guardian)
  2. ‘Dozens’ being investigated over Post Office scandal
  3. ‘Dozens of people’ being investigated over Post Office scandal, police chief reveals
  4. Oldest Post Office victim offered a third of payout

NotebookLM original audio link: https://notebooklm.google.com/notebook/ed04c18c-3a9e-447a-b14b-b99a14af9777/audio

I’ve split this into two parts (generated as such) and uploaded it to Spotify as Do Not Bend: The Post Office Horizon Scandal.

Decoding speech perception paper

I grabbed the arXiv HTML link for Decoding speech perception from non-invasive brain recordings and provided that directly to NotebookLM via the link option.

The podcast is a pretty good high-level overview. NotebookLM audio here: https://notebooklm.google.com/notebook/e1c7cda0-d065-427f-99e4-90df83066501/audio.

Podcast Workflow via Spotify

Downloaded audio can be split with Audacity (my . When splitting the track, you can export multiple WAVs (44.1 kHz at 16-bit PCM) in one pass via File > Export Audio and select Export Range: “Multiple files”.

Audacity Export Multiple Files

Upload to Spotify is easy via the Spotify for Creators Dashboard. The Spotify for Creators Terms and Conditions of Use are too onerously long.

I created a two-part podcast called Do Not Bend getting my cover art from Unsplash (this one), which has a permissive license (all images can be downloaded and used for free for commercial and non-commercial purposes with no permission needed).

Spotify for Creators Dashboard

The UI for upload to Spotify is straightforward.

For me the hosting was instantaneous.

You can use RSS to host trivially on other platforms like Apple Podcasts or Amazon Music. Adding an email allows you to verify on those other feeds.

Et voilà, we’re live 🎧

Addendum: I just wanted to see how this worked start-to-finish but of course, people have done this a bunch.

Graph View

Backlinks

  • No backlinks found