Earkind

Earkind is a pipeline that builds full AI podcast episodes from a text list of news and arXiv paper links, using tuned language model prompts, character voices, Azure text to speech and programmatic audio editing with Pydub, plus auto generated titles and timestamped show notes.

Earkind is an experimental podcast production pipeline that turns a plain text file of news items and a short list of arXiv paper links into a complete, produced podcast episode. The project combines language models, expressive neural text to speech and scripted audio editing to output something that sounds assembled rather than dictated: an intro, spoken segments, transitions, a sponsor read, an outro and a generated episode description with timestamps. The workflow is deliberately hands on. A .txt file holds the news the creator wants covered plus three recent arXiv URLs. Titles and abstracts are pulled from arXiv, and further detail is extracted from the raw PDF text through the ChatGPT API. Prompts are written and tuned per section (intro, outro, transitions, sponsor) and per story, using zero or one shot examples depending on how much creative latitude the section needs. Discussion is structured as a conversation between defined characters: Giovani Pete Tizzano, an overhyped but earnest tech enthusiast; Robert, a sarcastic analyst; and Belinda, a witty research specialist. Scripts are parsed for speaker markers, then voiced with Azure text to speech. Audio assembly is handled with Pydub. Jingles, sound effects and background music are mixed with the narration, volumes are adjusted automatically, and sections are overlaid and looped. Because the edit is programmatic, the tool also knows the timing of each segment, which is how the timestamped episode description is generated; titles and the overall blurb come from ChatGPT as well. Earkind suits podcasters who want a repeatable news or research show without booking a studio or co hosts, and technically curious builders who want to inspect and modify a full generation pipeline. The published GPT Reviews feed shows the intended tone: informal, character driven and a little over the top. It is not a hosted studio for uploading your own audio, and the project describes itself as rough, with code release and a deeper design writeup planned.