Every podcaster knows the feeling: the episode is done, the audio is clean — and now you have to find the 40 seconds worth clipping for Shorts, Reels, or TikTok. So you drag the progress bar, listen, drag again, listen again. An hour of audio can eat an afternoon.
The frustrating part isn't the editing. It's the finding. Most tools that promise to fix this either upload your audio to the cloud (privacy + subscription) or lean on AI transcription APIs that bill you per minute of audio. For a hobby podcast or a small creator channel, that's a non-starter.
The simple physics of highlights
Turns out there's a much more boring, much more reliable signal: people get louder when they say something worth clipping. A guest's best story, the host's hot take, the payoff of a long setup — they all register as short bursts of higher RMS energy against the surrounding conversation.
So instead of listening to the whole episode, I wrote a tool that does the listening for you:
- Reads a 16-bit PCM WAV file frame by frame (100 ms frames by default)
- Smooths the energy curve with a sliding window so coughs and one-off noises don't register
- Finds contiguous regions above the threshold and merges adjacent ones
- Ranks the candidates by energy strength and outputs the Top N highlights
The whole scan runs in about 30 seconds on a typical episode — and the output is a checklist you can act on immediately.
Output that's actually actionable
Here's what the report looks like:
## 片段 1: 12:34 - 13:05 (时长 31.0s, 强度 2.34x)
ffmpeg -ss 12:34 -to 13:05 -i episode.wav -c copy clip_01.mp3
Copy the ffmpeg command, paste it into a terminal, done. One command per clip. No editing timeline, no export dance, no cloud upload.
A few knobs for different content styles:
-
--frame-ms— finer frames for fast-paced content, coarser for calm interviews -
--window-sec— bigger smoothing window if your speaker ramps up slowly -
--min-duration— filter out coughs and one-word interjections -
--top N— get more or fewer candidates per episode
Why local matters
Your raw episode is often unreleased or contains client conversations. Sending it to a third-party service to get a clip list back is a real privacy decision. This tool never moves your audio anywhere — the WAV stays on your machine, the energy analysis happens in memory, and the only thing that leaves your terminal is the ffmpeg command you choose to run.
The full kit
The tool is a single-file Python CLI, pure standard library, zero dependencies — it runs anywhere Python 3.8+ runs. It's part of the AgentChip toolkit: a growing library of small, local-first tools for people who build and create. One-time purchase, free updates, instant download.
Your episode has 60 minutes of audio and maybe 2 minutes worth clipping. Now you can find them in half a minute.
Originally published on the AgentChip blog.
