ShortsAITurn a podcast into shorts without cropping out half the conversation
A two-person podcast is the hardest input an auto-clipper faces: a single centered crop shows the wrong person half the time, or an empty chair. ShortsAI detects a real two-shot and renders both speakers stacked in half-frames, so a reply never happens off screen.
The rest of the pipeline is the same as for any long video: word-level transcription, scene detection, Gemini scoring the 3 to 15 strongest moments, subtitles burned in, and direct publishing to TikTok, Instagram Reels and YouTube Shorts.
Cost is where podcasts punish credit-based tools: they bill the whole episode length before you see a clip. ShortsAI has no meter at all and no paid tier.
Why podcasts break naive clipping tools
Podcast video is a conversation, and conversations move. A tool that crops a fixed center column out of a wide two-shot shows whoever happens to sit in the middle, which is often nobody. A tool that follows one face loses the reaction shots that make clips work. Reviewers of the mainstream clippers consistently report exactly this: framing that needs manual correction on multi-person footage. It is not carelessness, it is that a single moving crop cannot show two people at once.
How the two-speaker layout works
ShortsAI detects when a scene is a genuine two-shot, meaning both faces are visible in the same frame for at least half of the sampled frames. That test matters: it is what separates a real side-by-side conversation from shot/countershot editing, where a naive split would show the same person twice. Confirmed two-shots render as a split layout, both speakers stacked in half-frames filling the 9:16 canvas. With speaker cutting enabled the clip instead hard-cuts to whoever is talking, with mouth activity normalised per speaker so that lighting and contrast differences do not hand the whole scene to one side of the table.
From episode to posted clips, step by step
- Paste the episode's YouTube link or upload the file. Podcasts of an hour or more are the normal case, not the limit.
- faster-whisper transcribes with word-level timestamps, and PySceneDetect maps the cuts.
- Google Gemini reads the transcript against the scene boundaries and returns the 3 to 15 segments that stand alone best, 15 to 60 seconds each.
- Each segment is reframed for its content: split layout for two-shots, face tracking for single speakers, screencast layout if the episode shares a screen.
- Subtitles are burned in from the word-level transcript, and finished clips post directly to TikTok, Instagram Reels and YouTube Shorts, or come back through the API.
What a full episode costs to clip
Credit-metered tools bill on the length of the video you import, not on the clips you keep. As of August 2026, a 60-minute episode costs 60 credits at Opus Clip or Vizard whether it yields 5 usable clips or 20, and a weekly show at that length runs past the entry plans of both. ShortsAI prices the other way around:
ShortsAI comes in two editions and they are priced very differently, so it is worth being precise. ShortsAI has no paid tier at all. Free at shortsai.app, with nothing to install. You paste your own Google Gemini API key (free tier: 1,500 requests a day) in Settings and it stays encrypted in your browser. No plans, no watermark, no usage cap, no account. The site is supported by ads. You bring your own Google Gemini API key, whose free tier covers 1,500 requests a day; the site is supported by ads.
What about audio-only podcasts?
ShortsAI clips video. If your show is audio-only, the pipeline has nothing to reframe, and tools that generate waveform audiograms serve that case better. The moment you record video, even a static two-camera setup, everything on this page applies.
Common questions
- How do I turn a podcast into clips for free?
- Paste your episode link into ShortsAI at shortsai.app with a free-tier Google Gemini API key: no watermark, no cap, no paid tier, no account.
- How does it handle two people talking?
- Scenes where both faces share the frame at least half the time render as a stacked split layout so both speakers stay visible. Optionally it hard-cuts to the active speaker instead, using per-speaker normalised mouth activity to decide who is talking.
- Does it work with hour-long episodes?
- Yes, long-form is the design case. Processing time scales with length: on CPU roughly 5 to 8 minutes of processing per 8 minutes of source, on an NVIDIA GPU about a tenth of that. The AI moment scoring reads the transcript rather than the raw video, so episode length does not degrade selection quality.
Sources
- Competitor per-minute credit billing checked 2026-08-04 on vendor pricing and help pages.
- Split-layout and speaker-cut implementation in the project source at shortsai.app.