How to compile every clip of one speaker from a long recording
Give Vizard Agent a long recording and say whose parts you want. Vizard Agent transcribes it with speaker labels, checks the frames to confirm the right person is on screen at those timestamps, and builds the compilation two ways: one that keeps the natural flow, and one that keeps only that speaker's words.
What do you type into Vizard Agent?
Name the person and the recording. Vizard Agent handles the speaker separation and the verification, so you don't need timestamps or a count of sections: it works both out from the transcript. Say how strict to be too. Natural flow keeps a little context around each answer; strict keeps only their words.
Prompt
What does Vizard Agent actually do?
Finding a speaker's parts is a transcript problem, cutting them cleanly is a seeking problem, and getting the audio to survive the cut is a third problem entirely. So Vizard Agent spends the first half finding the sections and the second half verifying the render. In the session behind this article it:
- Downloaded the recording from the link and generated a word-timed transcript with speaker labels.
- Extracted frames at the timestamps where that speaker was talking and checked the grid to confirm the right person was on screen.
- Analysed the exact word boundaries for that speaker and built both versions, natural flow and strict.
- Analysed both renders for cut cleanliness, found visual glitches and re-rendered with frame-accurate seeking.
- Checked the codec streams and the volume, found the audio track had come out silent, traced it to the order of the seek flags, and re-rendered both versions with the audio verified present and loud. The delivered cut ran 97.46 seconds at 1920x1080 and 25fps.
When does this not work well?
Speaker separation is good, not perfect, and a compilation is only as reliable as the labelling underneath it. Overlapping speech belongs to both speakers or neither, and off-camera speakers have no picture to check against. Strict cuts lose the question, which is what the natural-flow version is for.
Common questions
Does it know who is who? It separates the speakers and checks the frames at those timestamps to confirm the labelling matches the picture.
Will the audio be right? Vizard Agent measures the volume of the finished render rather than assuming the track came through.
Can it handle a two-hour recording? Yes. Length is not the constraint. Vizard Agent transcribes the whole thing and works from the timings.