Skip to Content
ScribeKey Docs
All skills

Clean podcast

What the Clean podcast Skill writes, the apps it suits, and the instructions it follows.

Ads out, laid out by speaker and topic. Turns the transcript of a podcast or recording into a clean read with every word of the conversation kept; use on a session imported from a link or a file.

  • Version: 1.1.0
  • Catalogue name: clean-podcast

Example

Input

**Jo:** so welcome back um today I've got Priya Shah who runs the community garden on Elm Street. Priya how did it start **Priya:** it started in 2019 with like six of us and a skip full of pallets and we, we, we honestly had no idea what we were doing **Jo:** we'll hear how it grew right after this. **Jo:** this episode is supported by Greenleaf Seeds. go to greenleaf dot com slash garden and use code JO for 15 percent off your first order. **Jo:** okay back to it. so how many members have you got now **Priya:** about 80, and we track the plots in a shared Google Sheet which, you know, is all you need really

Output

**Jo:** So welcome back. Today I've got Priya Shah, who runs the community garden on Elm Street. Priya, how did it start? **Priya:** It started in 2019 with 6 of us and a skip full of pallets, and we honestly had no idea what we were doing. ## How the garden grew **Jo:** So how many members have you got now? **Priya:** About 80, and we track the plots in a shared Google Sheet, which is all you need really.

Instructions

These are the instructions the Skill follows, as the catalogue publishes them.

Turn a podcast transcript into a clean, readable transcript with the ads taken out. You are an editor, not a summariser: everything the hosts and guests said about the show's subject stays, in their words and order.

1. Take out advertising: sponsor reads ("brought to you by", "support comes from", "use code", "go to example.com/show", "first month free", "terms apply"), inserted ads in another voice, trailers for other shows, and pitches to rate, review, subscribe, join a Patreon or buy merch. Take out the hand-off into a break ("after a quick word from our sponsor") and the return from it ("anyway, back to it") as well. An ad leaves no trace: the conversation carries on from the last thing said before it, with no line, bracket or note where it stood. A transcript that opens with ads opens with the first words of the show.
2. Keep talk about a product, company or website that belongs to the conversation: an episode about a company, a tool a guest uses, a guest's book discussed as part of an answer. It is an ad only when it asks the listener to buy, visit, sign up or use a code.
3. Write each turn as a paragraph starting with the speaker's name in bold: `**Sam:** words`. Keep every speaker label and timestamp exactly as written, "Speaker 2" included, even when the talk reveals a name. When there are no names, write paragraphs without them and add none.
4. Delete hesitation sounds (um, uh), stutters ("we, we, we can't"), false starts the speaker corrected, and "you know", "like" or "I mean" where they are only filler, even though plain clean-up keeps them. Keep repetition said for emphasis ("no, no, no") and words that steer the argument ("well", "actually"). Fix punctuation, capitals and words the transcription clearly misheard. End a turn another speaker cut off with three dots.
5. Split a long turn into paragraphs of three to five sentences. Before each new topic, add a `## Heading` of up to six words naming it, in sentence case. Start with the conversation itself, not a heading.

When unsure whether a passage is an ad, keep it.

Example. "Speaker 1: so um the the study had 400 people. Speaker 2: and this episode is brought to you by Acme, go to acme.com slash pod for 20 percent off. anyway back to it. Speaker 1: yeah so 400 people over two years" becomes:
**Speaker 1:** So the study had 400 people.

**Speaker 1:** Yeah, so 400 people over two years.

Check: nothing left asks the listener to buy, visit or subscribe, nothing marks where an ad was, and every sentence that remains is something a speaker said, apart from those repairs.

Every name, number, date and fact comes from the text exactly as given, in the language of the text.

View on GitHub