AI models are meant to be specialized for specific use cases! Take Google's NotebookLM, a RAG-based workspace to content-create and ask questions over specific document sources. One feature is dubbed "Audio Overviews" that transform the content into a podcast-style overview with two artificial co-hosts. I've long been a listener of Techmeme Ride Home with Brian Mccullough?as the host who is always experimenting with the latest in AI from ElevenLabs to now NotebookLM. He took today's episode and ran his notes through NotebookLM: https://lnkd.in/eGx3S9hU
While I do not have access to his notes and script, I made the following observations:
- The synthetic audio personas sounded extremely lifelike, there was no listening fatigue and I sincerely forgot I was listening to AI-generated voices about 2 minutes in.
- The generated podcast script was engaging and imbued the co-host personas with personalities who crack jokes and banter like two podcast hosts would. They ask questions, make commentary, converse, and delve into the topics at hand as well as use casual language, all of which made me chuckle and maintain my engagement.
As only speculation as an AI practitioner, Audio Overviews in theory uses a model by Google trained on tons of podcast audio and scripts to both mimic the podcast "style" of voice, both audio and text. This proves that highly use-case specific AI models can excel quite well!
#Techmeme #ridehome #NotebookLM #GenerativeAI
Ben & Jon are such thoughtful founders. It was a pleasure interviewing them and hearing about their plans for Ensis!