Skip to main content
Micaela, Zazu, and I were fascinated and stunned as we listened to David Levine’s AI production–a word-for-word, but radically reorganized overview of Zazu Dreams.
Both tantalized and totally disoriented, we are including this AI-generated conversation to publicize our book that amidst our many interwoven topics, explores techno-euphoric, cyber-addicted realities. Are we self-serving hypocrites or have we embodied the pluriverse?
Levine writes:
“This audio was generated using Google’s NotebookLM product, based on reading the entire text of the Zazu Dreams novel. NotebookLM is currently the best available technology for producing the most realistic sounding conversations using AI. From a linguistics perspective, the secret sauce is the use of significant paralinguistic training in the Large Language Model (LLM) behind the AI generating the speech, which inserts natural-sounding non-dysfunctional disfluencies, combined with the use of backchannel cues from the listener, to create a dialogue that is getting close to climbing out of the uncanny valley of AI-generated audio conversation. In other words, the speaker says things like, ‘Um, uh, like’ and so on, while the listener inserts sounds like ‘yeah’ and ‘wow’ and other non-language sounds that indicate the listener’s response to the speaker. And there is excellent prosody as well.
At the LLM level, Zazu Dreams was ‘read and comprehended’ by the AI using an AI technique called Retrieval Augmented Generation, which essentially adds an understanding of the novel to its existing understanding based on scanning, well, the entire web or something like that. The AI was then instructed to generate a script for two characters with personalities to discuss this novel. I manually mixed in some audio at the beginning that is open source and licensed. I should add that I didn’t even touch on the most fascinating part of all this, which is the fact that this is generated by AI and what that means. I could write an entire book on that topic.”

Leave a Reply