How vodformer finds the moments in a 4-hour stream
A four-hour stream holds a handful of moments worth posting. Finding them by hand means scrubbing the whole thing at double speed, listening for chat to go wild. vodformer does that pass for you, then hands you the picks with the reasons it chose them. This post explains what it looks at and where it still falls short.
What a stream gives us to work with
A stream VOD carries more than the video. vodformer reads several things from it:
- The transcript. The whole stream is transcribed with word-level timestamps. That lets us find stories and hot takes even when nothing loud happens, and it lets every cut land on a word boundary.
- The audio. We measure how loud the streamer is compared with the last couple of minutes, and an audio model listens for laughter, screams and cheering.
- The chat replay. On Twitch, and on Kick when the history is still there, we count bursts of messages by kind: laughing, hype, shock, and people typing “clip it”.
- Viewer clips. On Twitch, viewers often clip the good bits during the stream. Twitch records where each clip sits in the VOD, so a cluster of viewer clips is a strong hint.
- The picture. An AI model watches a low-resolution copy of the stream and suggests moments that are visual, like a play in a game, which the other signals can miss.
An uploaded file has no chat and no viewer clips, so for uploads the transcript, the audio and the picture do the work.
From signals to picks
Each signal produces candidate moments. A language model also reads the transcript in chunks and adds moments that the signals don’t catch, like a good story told in a calm voice.
Then the candidates are ranked together. The model that does this answers with sentence numbers from the transcript, not timestamps, so a pick can only start and end where a sentence does. The edges are then moved into the nearest pause, so a clip doesn’t start mid-word or cut someone off.
Each pick comes with chips that say why it was chosen: “Chat: laugh”, “Laughter”, “3 viewer clips”, a quoted line, or where the pick came from. If a pick only has one weak reason, you can skip it without watching.
The best few picks are drafted as vertical shorts straight away, with a layout and captions, so you can open one and start editing.
Where it still goes wrong
The clipper is new, and it misses. The common failures:
- Quiet channels. With little chat and no viewer clips, the picks lean on the transcript and the audio, and the ranking gets less sure.
- Context. A moment that’s only funny because of something said ten minutes earlier looks flat on its own.
- Edges. A pick can start a beat too late or end before the punchline.
That last one is why the editor matters as much as the clipper. In vodformer you can drag a clip’s start and end past where the AI cut it, delete words from the transcript to tighten it, or stitch two moments into one video. A wrong pick costs you a few seconds of fixing, not a re-export from another tool.
What we learn from testers
During the alpha we record what testers do with each pick: kept, trimmed, deleted, exported. That tells us which signals to trust for which kind of stream, and it is how the picks get better. If a pick is badly wrong, the Feedback button in the editor goes straight to us.
vodformer is in a closed alpha. Join the waitlist to try it.