Skip to content

Finding clip moments in a transcript

Reading and correcting a machine transcript, and asking it for places worth a look - which is a text search, not a judgement.

Updated

The transcripts page holds machine transcripts of your renders and uploads, lets you correct low-confidence segments, and can point at places in the words that might be worth clipping. It is what a model produced, not a checked record, and its standing is shown so you know which is which.

"Look for clip moments" matches question marks, figures, list cues and speaker changes. Those are patterns in the words, not an assessment of what was said, and the results are deliberately not ranked.

BEFORE YOU START

  • A transcript. Transcripts arrive from a transcription render, and this build has no transcription integration at all - so unless transcripts reached your organisation some other way, this page will stand empty.
  • The "view" or "generate" permission to read; "generate" or "administer" to correct and to mark reviewed; "administer" to remove a speaker or delete a transcript.
  • Nothing else. Nothing on this page costs credits.

STEPS

  1. Open /app/transcripts.
  2. Choose "Read it" on a transcript. Segments are shown with their timecode and the model's confidence.
  3. Segments the model was unsure about carry a "check this" chip. For each, either choose "These words are right" or choose "Fix them", correct the text and choose "Save correction".
  4. When you have been through it, choose "I have read this transcript and it is accurate". That changes its standing, which is recorded on every subtitle you export.
  5. Choose "Look for clip moments" to see candidates. Each shows a timecode range, its length in seconds, the text and why it was flagged.
  6. Choose "Cut this moment" on a candidate to carry the recording, the transcript and both timecodes into the clips page. It appears only when the transcript records which recording it is of, which is not always the case.
  7. Download subtitles as SRT or VTT from the list if you need them.

WHAT YOU SHOULD SEE

A list of candidates under "Places worth a look", with the method note printed above them, or a plain statement that nothing matched. Nothing matching does not mean there is nothing worth clipping; it means no question, figure, list cue or speaker change turned up.

WHAT THIS WILL NOT DO

  • It does not judge quality. Confidence measures how sure a model was, not whether it was right, and the candidates are not ordered because an order would imply a ranking the signals cannot support.
  • It does not cut a clip itself. Cutting one is a separate capability with its own page - see "Cut this moment" above, and the guide below.
  • It cannot remove one person from an undiarised recording. With no speaker labels there is no way to tell whose words are whose, so the only option is deleting the whole transcript.
  • It does not check the transcript for you. Marking one reviewed is a statement you are making, and it travels with every export.

IF IT DOES NOT WORK

  • "No transcripts yet. Transcriptions appear here once a transcription render finishes." is the standing state on a build with no transcription integration.
  • "Transcript not found." or "Segment not found in this transcript." means it belongs to another organisation or has been removed.
  • "Subtitle format must be srt or vtt."
  • If the speaker removal panel says the recording has no speaker labels, deleting the whole transcript is the only way to honour a removal request.

COMMON QUESTIONS

Why does a downloaded subtitle file matter whether the transcript was reviewed?

Because every export carries that standing in a header. A machine transcript nobody checked should not be presented as a checked one.

Can I clip a moment straight from here?

Yes - "Cut this moment" carries the recording, the transcript and both timecodes into the clips page, where you choose the platforms and the words. Note that cutting a clip is not rendering it: the render that turns a clip into a video needs a compositor, which no deployment has active.

  • Cutting a clip from a recording

    Turning a stretch of a long recording into a short, captioned video for each platform - and reading what a narrower shape costs before you commit to it.

  • Repurposing a source into several drafts

    Turning one article, page, document or recording into posts, captions, threads and hooks, with a record of where each draft came from.

  • Following a render

    The renders page, what each render state means, and the rule that a render is charged only when a provider produced something.

  • Rendering a composition

    What submitting a cut version does, why every deployment currently answers "no provider active", and what that costs.