Phase I: NotebookLM

These are experimental podcasts. They have been developed using AI technology.

Documents were entered in NotebookLM, an experimental “Note Taking and Research Assistant” powered by Google’s Gemini 1.5 Pro Large Language Model. Each podcast is an AI-created dialog using the documents as source materials. Some prompting is done to set the general direction of the discussion.

There is some additional material added in some discussions based on the information in the LLM obtained during training.

It’s important to note that all the podcasts are exchanges between two people.

Japan Travel Adventures

This podcast is based on the text of a travel narrative. The trip took place in the Fall, 2023. We visited Tokyo and Kyoto and did some day trips to other nearby cities.

Document Link (Click on ISSUU link):

• Japan Fall 2023 (ISSUU link)

Using an LLM to Create R code

A document was created to show how simple statements can create R code that will input and analyze data. The target audience is researchers in the general field of ethnobiology, although the technology applies to a much wider field of science.

Document (Click on Quarto Link):

• Using an LLM to Create R Code (Quarto Link)

The Species Problem

This is an experimental test to see how well the LLM (Gemini 1.5 Pro) handles the merging of two documents that are related but very different. One is the start of Chapter 13 (Systematics and Taxonomy: Testing evolutionary hypotheses) from the biology textbook by Robert Wallace & Jack King (1981) Biology: The science of life. The other text used as a source is Chapter 13 (The Rule of Five) from Stephen Jay Gould (1985) The Flamingo’s Smile.

This podcast is just about the Species Problem (Wallace & King textbook).

The second podcast combines the two Source documents.

Classic Papers

There are a series of publications that, when they were introduced, made quite a splash. It’s fun to revisit these documents, now with introductions from the folks at NotebookLM.

George Miller’s The Magical Number Seven, Plus or Minus Two (1956) in The Psychology Review tells us a lot about how humans handle data.

John Platt’s Strong Inference (1964) in Science describes how the methodologies in some branches of science lead to more rapid progress.

Jared Diamond’s The Ethnobiologist’s Dilemma (1989) in Natural History reminds us that it’s best to have some expertise in a field before asking other people questions.

Document Links (Click name for access):

• Magical Number Seven

• Strong Inference

• The Ethnobiologist’s Dilemma

Podcasts:

The Gallery Catalog

This podcast was built by combining several AI technologies. The core document is a “made up” gallery catalog. The style of the catalog attempts to emulate a real gallery catalog. There are biographical statements by the artists, an image of each artist, and several examples of each artist’s works. These elements were created using text-to-image software, AI-assisted text editing, and manual photo touchup.

The podcast software inside NotebookLM can’t do image recognition. Instead, each of the artworks was sent to OpenAI’s ChatGPT engine to get a comprehensive image description. The same process was used for the images of each artist and the gallery owner. Files with these descriptions were provided to NotebookLM.

Two additional text files were supplied to guide NotebookLM. One file describes the role of each of the files. An account of the process used to create the catalog is contained in the other file.

Document Link (Click name for access):

• Gallery Catalog (ISSUU link)

Phase II: Beyond NotebookLM

The experiences using NotebookLM were positive. However, there were some shortcomings of the first-generation implementation of NotebookLM. The next experiment probed the use of ElevenLabs Voiceover Studio (beta).

A NotebookLM Critique

The following podcast was created from a script. Each line of the script had a “speaker” and a “line”, all arranged as a CSV-format file. The file was input to Voiceover Studio, speakers were assigned voices, audio blocks were generated and the location of each block was adjusted on the timeline.

If you didn’t notice, this review podcast (above) uses all the features cited as shortcomings in NotebookLM podcasts: more than two speakers, editing and scripting.

An Alternative AI

Anthropic’s Claude was asked to generate a conversation between two people about Darwin’s book “Voyage of the Beagle.” One speaker would be a biology teacher and the other a specialist on travel literature. The output was requested to be in the CSV format used by Voiceover Studio. The results were processed the same as the NotebookLM Podcast Review.

Here are two versions of the podcast discussing the book. Each was generated using a slightly different prompt to Claude. The podcast audios are presented with very minor editing. Remember, editing is a useful feature when using the ElevenLabs approach.

Phase 3: New Tools

A Narrated Video

Anthropic’s Claude continues here as the AI engine to generate conversation texts.

The transcription of the text into audio is done with ElevenLabs project capability. This turns out to be a more efficient way of adding voices to the multi-speaker text. As shown earlier, this also provides an opportunity for editing and adding another cast member.

CapCut is used to combine images with the audio track. The video is uploaded to YouTube, a service that is invaluable.

Gallery Catalog (part 1): (YouTube link 4:13)