Streaming

Turn a stream VOD into finished videos

A four-hour stream is not one video. It is three or four, with a lot of nothing in between. The highlights finder reads the whole recording, works out which stretches are actually a story someone could watch start to finish, and builds each one as its own sequence.

What it is looking for

Not the loudest parts. Loud is easy to find and it is not the same as watchable: a stretch where you shout at a jump scare for four seconds is not a video. Four signals decide, and the first one carries the most weight.

• Topic cohesion. Does this stretch keep returning to one subject across its whole length? This is what separates a publishable cut from a merely eventful one. A viewer needs to be able to follow what is happening.
• Sustained reactions. Excitement spread across the block, not one spike. A viewer game or a challenge has energy throughout; a monologue with one laugh does not.
• Audience interaction. How much of the talk is aimed at chat: second-person address, reading names, taking requests.
• Speech continuity. A block that is mostly dead air is not an activity, however exciting its few moments are.

Step by step

1. Put the VOD on a timeline with the dialogue on an audio track. Open Um Out from Window → UXP Plugins and choose Find stream highlights.
2. Set the length range. The default is 15 to 30 minutes; the slider goes from 10 to 40. This is the length of each finished video, not of the stream.
3. Choose how many cuts to look for. Three is the default. Asking for more does not invent them: if the stream only contains two distinct activities, you get two.
4. Analyze. On a long VOD this takes a while, because it transcribes the whole recording. It runs on your machine.
5. Read what came back. Each proposal has a length, a plain summary of what happens in it, and title ideas. Play across the timeline bar to check before committing.
6. Approve the ones you want. Each becomes its own sequence in your project bin, with the dead air already trimmed out.

Why it does not offer the same video twice

A long stream returns to the same activity over and over, and two stretches an hour apart can be the same bit. Early versions of this tool would offer four proposals that turned out to be one video listed four times, which is worse than offering one: it fills every slot with a thing you have already seen and hides everything else that happened.

So it compares what was actually said in each stretch, weighting words by how rarely they appear in the rest of the recording. On a Minecraft stream the word “minecraft” is in every candidate and tells you nothing; what is left after you set those aside is what genuinely differs. Two windows that share most of that vocabulary are one video, and only one of them is offered.

Finding more cuts

Press Find different cuts on the results screen. Um Out remembers what it has already shown you and looks past it, so this gives you the next-best stretches rather than the same ones again. When there is nothing left it says so plainly instead of returning an empty screen.

Titles

Title ideas are starting points, not final copy. They are built as a claim about something you did, optionally qualified by the rule that made it interesting, because that is the shape this genre actually uses: “I Bought Every Crate On The SMP”, “Gorilla Tag, But The Lava Keeps Rising”.

Nothing in a title asserts an outcome that the recording does not support. Um Out reads a transcript; it does not watch the video, so it will not tell you the challenge went wrong when it has no way of knowing.

With Um Out Plus, or with your own API key, a model reads each segment and writes the titles and summary instead. Without either, the same job is done offline from your own words, and the panel says which one you are getting.

What gets trimmed as it builds

Two things, both optional and both shown to you first:

• Silences. On by default. The finished cut is tighter than the stretch it came from.
• Off-topic detours. Off by default, because it deletes footage you did not individually approve. With it on, a sustained wander away from the segment’s own subject is dropped and the cut joins to the next on-topic part.

When it finds nothing

It tells you why rather than showing an empty screen. The usual reasons:

• The recording is shorter than the minimum length. A 12-minute recording cannot contain a 15-minute cut. Lower the range.
• No stretch held one subject long enough. Common on a pure just-chatting stream, where the subject changes every few minutes by design. Lower the minimum length and try again.
• Everything left has already been shown. You have pressed “find different cuts” until the stream ran out.

What it needs

Transcription, which means the speech model has to be installed. The highlights finder itself is available on every plan and the search runs on your machine; naming the segments, writing the summaries, and building each cut with the trimming applied are part of Um Out Plus.