Podcast Transcript TXT Skill

Turn any podcast source into clean .txt — YouTube, episode webpages, Xiaoyuzhou, Apple Podcasts, X links, official transcript files, or just a title.

Philosophy

Audio locks you in. Text sets you free.

You can't annotate a voice memo. You can't skim a conversation, or put two interviews side by side and find where they contradict, or jump straight to the moment you half-remember from three weeks ago. The medium shapes what thinking is even possible.

A transcript doesn't replace the listening. It opens a different kind of access — one where you move at your own pace, mark what matters, combine fragments across sources. That's not a small thing.

This tool exists to make sure the text exists. Locally. Without sending your audio somewhere else first.

@一龙小包子 · 关心 AI，更关心人类

Output Quality

Every run produces a .meta.json alongside the .txt — it tells you exactly how the transcript was sourced. Official transcripts and subtitles are clean enough to read immediately. ASR output is a working draft: run one LLM proofreading pass before publishing. The tool always tells you which is which.

Recommended User Prompt (Copy-Paste)

Use this when handing the skill to end users:

If audio fallback is needed, choose ASR model:
- small (default): faster, lighter, best for first draft
- medium: slower, larger, usually better on names/terms

Output will include both:
1) ` - <title>.txt` or `<title>.txt`
2) matching `.meta.json` (resolver + quality + attempts for debugging)

Important: this transcript is a draft. Run one strong-LLM proofreading pass before publishing.

This project is designed for practical reliability:

• Prefer official transcript sources when available.
• Reuse visible page text or show notes before running heavy ASR when a host exposes meaningful text but no ready transcript.
• Fallback to YouTube subtitles when needed, then local YouTube-audio ASR if subtitles are unavailable or unusable.
• If no transcript/page text/captions are available, fallback to local ASR (faster-whisper, selectable small|medium, default small).
• Emit machine-readable diagnostics for every run.

TL;DR

git clone https://github.com/KingJing1/podcast-transcript-txt-skill.git
cd podcast-transcript-txt-skill
python3 -m pip install -r requirements.txt
python3 scripts/podcast_transcript_txt.py --doctor
python3 scripts/podcast_transcript_txt.py \
  --input "https://www.youtube.com/watch?v=n1E9IZfvGMA" \
  --out-dir "/tmp/transcripts"

Output:

• - <title>.txt when podcast name is already available
• otherwise <title>.txt
• matching .meta.json

Agent Quick Start

If you want another agent to use this repo directly from a GitHub link, this is the shortest reliable flow:

git clone https://github.com/KingJing1/podcast-transcript-txt-skill.git
cd podcast-transcript-txt-skill
python3 -m pip install -r requirements.txt
python3 scripts/podcast_transcript_txt.py --doctor
python3 scripts/podcast_transcript_txt.py \
  --input "<episode-link-or-title>" \
  --out-dir "/tmp/transcripts"

What --doctor means:

• Exit 0: this machine is ready for direct agent use, including ASR fallback.
• Exit 1: install or fix the missing runtime before asking the agent to transcribe.

Minimum pass condition for "directly usable":

• Python 3.9+
• yt-dlp
• faster-whisper
• writable model cache root

Features