Earkind is an AI-generated podcast project designed to provide non-boring audio content on AI topics. Its first show, "GPT Reviews," is a daily mixed-format program covering AI news commentary, attempted jokes, and deep dives into research papers. The show is presented by host Giovani Pete Tizzano (an overexcited AI follower) along with collaborators Robert (a frequently unimpressed analyst), Olivia (an extremely online internet surfer), and Belinda (a witty research expert).
The project operates by combining language models, neural expressive text-to-speech, and programmatic audio editing into a pipeline that can automatically generate complete podcast episodes and descriptions based on a list of news items and research papers. The specific process is as follows:
-
Input a .txt file containing several plain-text news items selected by the user as well as the URLs of 3 recent arXiv papers. Fetch the titles and abstracts from arXiv, and use the ChatGPT API to extract additional information from the original PDF text.
-
Create and adjust system and user prompts for each segment (introduction, outro, transitions, sponsor) and subsection (each news item/paper). Depending on the complexity of the prompt, use either 0-shot or 1-shot approaches. The content discussion is organized in the form of a dialogue between two characters. The character definitions include: host Giovani Pete Tizzano (an overexcited, annoying but sincere tech bro), Robert (a sarcastic and unimpressed analyst), and Belinda (a witty research expert who reads all the papers).
-
After creating scripts for each segment of the show and parsing them accordingly (for example, for speaker markers), enter the "recording phase." Use Azure TTS for speech synthesis.
-
After all recording is complete, proceed to editing. Use Pydub to combine narration and other audio, automatically adjust volume, overlay and loop segments, and so on. Various short music pieces, sound effects, and background music are also created.
-
Finally, automatically generate a timestamped podcast description (obtained from Pydub while editing the podcast), an overall description, and a title, which are also generated by ChatGPT.
The project is currently in its early stages, and the author welcomes feedback and plans to make the code public and release a more in-depth explanation of the design and ideas if there is interest.
