Skip to content

Native Studio filler-word removal using local word-timed captions #2960

Open

Description

@thesammykins

Description

Add native filler-word removal to the GPUI Studio editor so a recording can be cleaned up without exporting it to another editor or running an external script.

A local prototype used Cap's word-timed captions to remove whole “um” and “uh” spans by changing the editable clip timeline. It preserved the original media, camera and cursor tracks; the resulting project reopened and exported successfully, and the cuts were accepted after listening.

Proposed behavior

  • Provide a Remove Filler Words action alongside the native Captions controls.
  • Generate local word-timed captions automatically when they are missing, using the selected downloaded model.
  • Start with explicit English support and whole um, uh, uhm, er and erm hesitation tokens, including drawn-out forms; preserve lexical substrings, acronyms and context-dependent words such as “like”.
  • Apply editable timeline cuts with the existing ripple-edit and caption-projection helpers. Preserve source media and the source transcript.
  • Group cleanup into one undo step. Repeating cleanup should not cut the same source span again.
  • Keep the resulting clips fully editable: recover a boundary using trim handles, or split and delete a missed hesitation with the normal Studio tools.
  • Preserve natural pauses around cuts. Where recognizer word ends extend to the next onset, use the audio to retain a detected pause and audible release; retain existing timings when speech cannot be separated from noise.
  • Report skipped words near transitions/held frames and reject invalid or overlapping timestamps without partially changing the project.
  • Keep caption visibility as selected by the user; generating a transcript for cleanup should not implicitly turn on visible subtitles.

This is transcript-based cleanup. Playback review is still needed because recognition and word boundaries can be imperfect. An optional cleanup preference for new recordings can be considered after the native action and timing behavior are verified.

Additional Context

Activity

  1. linear commented on Oct 23, 2026

    @linear
  2. added a commit that references this issue on Oct 23, 2026
    cb65c22
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions