Read Podcast

English · 简体中文

Read the podcasts you don't have time to re-listen to. Read Podcast turns the episodes you pick into a three-minute brief plus the full conversation, edited like a magazine interview — speakers named, sections titled, timestamps kept — and saves every piece as Markdown in storage you own.

Live demo → read-podcast.xielevi.cc · How it works · Quick start · How it compares

Why

You finish a two-hour conversation that was worth every minute. A week later you want the one argument that made it worth it, and the only way back is to scrub through audio. Show notes list topics; AI recaps squeeze two hours into a few bullet points. The conclusions survive, the reasoning and examples do not.

Read Podcast keeps the conversation itself. It transcribes the episode you pick, then edits it the way a magazine edits an interview: verbal repetition goes, the arguments, evidence and back-and-forth stay. You get an article you can skim in three minutes, read in full, search, quote and keep.

Who it's for

• Heavy listeners of long-form interview shows — business, science, history, tech — who want to return to an argument without re-listening.
• Researchers, writers and students who quote podcasts and need accurate, attributable text.
• Note-takers who keep a Markdown vault (Obsidian, a Git repository) and want episodes in it as real documents.
• Self-hosters who prefer their own deployment and their own model keys to another subscription.

It is a single-owner workspace for personal or small private deployments, not a multi-user service.

How it works

1. Subscribe — search Apple Podcasts or paste any RSS feed. You can also upload your own audio.
2. Transcribe — pick an episode and press Generate. Transcription runs on your own Whisper service (MLX on Apple Silicon, Faster-Whisper in Docker) or a cloud provider.
3. Refine — any OpenAI-compatible LLM edits the transcript into a long-form manuscript, with prompts written for the episode's language: English episodes get English instructions and come back as natural English, Chinese episodes as Chinese.
4. Read — open the three-minute brief, skim the one-line summary under each section, then read the full text in a reader built for long articles, on desktop or phone. Every manuscript is also a Markdown file you own.

Before and after

What speech recognition gives you (illustrative — typical unedited ASR output, written for this README):

so yeah um cities already measure noise right there's maps limits fines and and why
isn't that enough well because a noise map it it averages everything into one number
so a street with like a steady sixty decibels of traffic and a street with forty five
and a car alarm every every ten minutes they can look kind of similar on the map

What Read Podcast publishes (from the demo episode in the screenshots — an original, fictional conversation written for this README):

## 02 | Measuring noise versus listening to it

**Nora Hale**: Cities already measure noise. There are maps, limits, fines. Why isn't that enough?

**Sam Okafor**: Because a noise map averages everything into one number. A street with a steady
60 decibels of traffic and a street with 45 decibels and a car alarm every ten minutes can look
similar on the map, but they feel completely different to live on.

Length is checked, not hoped for: by default a draft shorter than 70% of the transcript fails the quality gate and is not published. On the live demo, a 63-minute episode (忽左忽右 #502) became a manuscript of about 17,000 characters: a timestamped outline, six titled sections and both speakers named.

Brief first, full text when you want it

Every manuscript opens with a three-minute brief: the main conclusion, key points, verbatim quotes and who should read on. Each section starts with a one-line summary, so you can skim the structure and dive into the parts that matter. On a phone it reads the same way.

Sample manuscripts

Read full results on the live demo, the maintainer's own instance, open read-only to visitors (generating new manuscripts requires your own deployment). The demo library is Chinese-language; switch the interface to English with the 中 / EN button:

• 宇宙超级工程是如何垮掉的：刘怡谈「挑战者号」事故四十周年 — 忽左忽右 · 63 min · two speakers, 6 sections
• 袁长庚×胡安焉：脏活与狗屁工作，谁在工位追问意义？ — 上海书评 · 80 min · three speakers, 11 sections
• 对话经济学家朱宁：1929之后，我们学会避免下一场经济危机了吗？ — 知行小酒馆 · 73 min · two speakers, 12 sections

Screenshots come from a local instance subscribed to public English podcast feeds (show names, artwork and episode titles from their RSS). The readable articles in the screenshots are placeholders, and the featured episode is fictional; no real transcript is reproduced.

以上为 README 的开头部分。其余内容请阅读原文。 查看完整 README ↗