How it works
What happens between the link you paste and the episode in your podcast app.
From a link to an episode
Reading the page
We fetch the page and take out the article itself: its text and footnotes, and the shape of the harder parts, such as transcripts, equations and embedded posts.
Who says what
Each quotation is matched to the person who said it, from the evidence beside it. Where that isn't clear, the quote is read in a neutral quote voice rather than a guess.
Reading for the ear
Numbers, dates, abbreviations and symbols are written out the way a person would say them.
The cast
The narrator and each person quoted get a voice of their own, kept for the whole episode.
Recording
Each passage is spoken by a text-to-speech model, and every take is checked: its length, its waveform, its pitch against the voice's own sample and, where transcription is available, its words. A take that fails is recorded again, a limited number of times.
Finishing
Speech levels are balanced and peaks limited, chapters are set at the article's section breaks, and the episode is published to your private podcast feed.
Images and code
Images get a spoken description, and code blocks a spoken account of what they do. Both are on by default, and one switch turns them off for an episode.
Languages
The article's language is detected and followed throughout: the voices, the reading rules and the checks.
What the checks can and can't tell
The checks help choose the best take. They don't certify who said what. Where transcription isn't available, a take is recorded as unchecked.
Every change to the service is on What's new. Currently v3.4.0.