Why Do AI Captions Show Gibberish on Silent Shots?

Updated 2026-10-27·3 min read·Dialogue & audio

Why it happens

A transcription model is trained to turn audio into text. Hand it a clip that's only room tone, traffic or music and it still tries. Whisper in particular tends to "hear" one of the phrases it saw constantly in training, like the sign-off at the end of a video.

We saw this on a phone-insert shot: a close-up of a text message, scored with ambience and no voice. The burned-in captions read, in Korean, "subscribe and like, it means a lot to me." The caption font had no Korean characters, so the whole line came out as a row of empty boxes across the frame. It looked like a font bug, but the transcriber had made up a sentence.

Two details are worth knowing:

• Priming doesn't save you. Telling the transcriber which line to expect didn't stop it inventing one.
• A written line isn't a spoken one. A shot can have a line of voiceover written on it that was never generated. The script says there's speech; the audio has none. Transcribe that clip and you get a hallucination.

How to fix it

1. Read the caption text before you touch fonts. Open the subtitle or caption file. If the boxes are a sentence you never wrote, it's a content problem. The number of boxes usually matches the invented sentence, not your line.
2. Only transcribe clips that contain speech. Before captioning a shot, check whether there is a voice in the audio itself. A line in the script doesn't count.
3. Caption silent shots from the text, or turn captions off. If the line is a text message or an on-screen sign, use the written text directly. If there's nothing to read, leave the caption off.
4. Make sure the voiceover really exists. If a shot should have narration, confirm it was generated and mixed in before captioning.
5. Lip-synced shots are the exception. When the voice is baked into the clip's own audio, transcribing that clip is correct, because the speech really is there.

Example

A thriller episode cuts to a close-up of a phone: "Where are you?" on the screen, a clock ticking under it, no voice. Run every shot through auto-captions and that insert gets a made-up line. Caption it from the text instead ("Where are you?") or leave it bare, since the audience can read the phone.

The reverse case: a narrated shot where the voiceover was written but never generated. Captions pulled from the audio will invent something. Generate the voice first, then caption.

How ShowMaker handles it

In ShowMaker every line lives in the script and is voiced by that character's fixed voice, so you can see which shots actually carry speech before anything is captioned. Captions and subtitles are added in the editor, in several styles. If you're fighting captions that drift instead of captions that invent, see why AI captions go out of sync, or try the AI video subtitle generator.