If automatic captions display words where nobody speaks, listen to that interval in the source audio before editing the transcript. Speech recognition can produce plausible text that was never said. The official Whisper model card documents this limitation. It does not mean every captioning service uses Whisper, and it does not justify assuming that every unfamiliar word is a hallucination.
Establish what is actually audible
Inspect the original audio with enough context before and after the caption. Is there quiet speech, a lyric, a distant television, a breath, or only music and room tone? Check whether your editor is playing the intended track and whether a muted dialogue track exists elsewhere. A waveform can help locate sound, but it cannot establish the words by itself.
Use comfortable, consistent monitoring rather than pushing volume to an uncomfortable level. If the source remains unclear, mark it for review instead of confidently substituting a sentence. When another person knows the recording, ask them to listen without first showing the suspicious text; the displayed phrase can influence what a listener expects to hear.
Keep a no-speech audit record
| Interval | Audible evidence | Generated caption | Editorial action |
|---|---|---|---|
| Intro with music | No intelligible speech after review | A complete invented sentence | Remove the false speech cue |
| Quiet breath | Breath, no word established | A short guessed word | Review context and remove if unsupported |
| Distant audible voice | Real speech, wording uncertain | A confident sentence | Transcribe only what can be supported |
These are hypothetical review cases, not measured results. Enter the real cue times and file name in your own record. Pay attention to openings, long pauses, music-only sections, and the tail after the last spoken line, but do not assume a fixed silence threshold will solve every recording.
Delete the false words without shortening the scene
Removing an incorrect caption should normally remove the subtitle cue, not the underlying picture interval. If the editing tool links transcript deletion to timeline deletion, choose the operation carefully and check the duration afterward. A pause can be intentional storytelling even when there are no words to display.
Non-speech captions serve a different purpose. An audible door slam or meaningful music may warrant a sound description for accessibility, depending on the project's caption policy. Do not preserve invented dialogue by relabeling it as an accessibility cue, and do not remove real sound information simply because it is not spoken language.
Verify the exported captions
Play the exported video with its final caption file or burned-in text enabled. Revisit every corrected interval and confirm that valid neighboring speech still has the correct start and end times. Check that a deleted sentence has not survived in another subtitle track or export preset.
For an Imagild film delivery, treat an automatic transcript as an editable draft. Keep the original audio available for verification and record unresolved wording for a human review. This workflow addresses unsupported speech text; it is separate from fixing a wrong input track, translating dialogue, or describing genuine non-verbal sounds.