How to trust the burnt-in captions over the audio when a name is misheard
Tell Vizard Agent that the video already has the words on screen and to read those instead. Speech recognition turns unfamiliar brand names into ordinary words that sound similar, and once that is in the transcript every graphic and every b-roll choice built from it is about something else entirely. Reading the burnt-in captions off the frames settles it.
What do you type into Vizard Agent?
Point Vizard Agent at the disagreement rather than asking it for another transcription. Running speech recognition again produces exactly the same mishearing, because the audio has not changed and neither has the set of words the recogniser expects to find in it. Name the word it got wrong.
Prompt
What does Vizard Agent actually do?
A transcript is a guess about sounds. A caption burnt into the picture is a decision somebody made about the words themselves, and where the two disagree about a name, the one on screen is almost always right. So Vizard Agent reads the text rather than inferring it. In the session behind this article it:
- Transcribed the speech to find the moments and built graphics and b-roll against that first reading, which illustrated a word the speaker never used.
- Was told by the user that the name was wrong, and instead of listening harder, read the captions directly off the video's frames.
- Took the key moments from the on-screen text rather than from the transcript.
- Rebuilt the graphics against the corrected wording and re-placed the b-roll at the right moments.
- Checked each insert against the frames it sits on, not only against the transcript's timings.
When does this not work well?
This only helps where text exists. A video with no captions, captions that are themselves wrong, or a name that is spoken but never written leaves you back with the audio and a spelling you have to supply yourself. Stylised or animated text is harder to read reliably, and a fast caption may never sit still on a frame.
Common questions
Why does transcription get names wrong? Because it fits sounds to words it expects, and invented names are not among them.
What if there are no captions? Then tell Vizard Agent the spelling and it uses that.
Will it correct the spelling everywhere? Yes. Vizard Agent applies it to graphics that were already built.