A free, open-source Python tool

Podcastify

Turns an e-book or article into a two-host podcast. In LLM mode it reads the whole source, writes a conversation in which every factual line cites the sentences it rests on, checks each line against them, and renders an MP3 with chapters.

From the eval run (eval/run_eval.py, the same pipeline as --mode llm --minutes 6) on the whole of Alice's Adventures in Wonderland (Project Gutenberg, public domain), 2026-10-29. S7 is the seventh sentence of the source; every claim line was checked against the sentences it cites. The write-up has the full episode, a transcript that opens each line onto its source, and the eval.

Two modes

rule

The default, with no model. The narrator reads the source in three-sentence segments and the co-host interjects from a fixed pool. Faithful to the text, and a reading rather than a conversation.

llm

Maps the whole source to cited key points, picks an outline for the target length, writes the host and co-host script with citations, and checks every line; lines that fail are rewritten once or dropped. Runs through the local claude -p CLI on a Claude subscription, or the Anthropic API when ANTHROPIC_API_KEY is set.

What the eval found

On four sources (a novel, an essay, a nonfiction chapter and a Wikipedia article), six-minute LLM episodes covered 7.4 of 12 frozen key points on average, against 4.4 for rule mode reading the same number of words. A blind judge preferred the LLM script on all four. The grounding check failed 9 of 151 claim lines, but a second checker found as many unsupported lines after the repair as before it, so what the check mainly adds is a record a reader can verify line by line.

Run it

pip install -r requirements.txt
python podcastify.py book.epub                  # rule mode, no LLM
python podcastify.py book.epub --mode llm       # grounded conversation
python podcastify.py book.epub --mode llm --minutes 12 -o book.m4b

ffmpeg and ffprobe are required. Voices default to en-US-AndrewNeural and en-US-AvaNeural; --voice-a and --voice-b take any edge-tts voice.