Allow deletion/removal of specific segments from transcripts

[hyper-charlie]

Use Case

Users need the ability to remove specific segments of speech from their transcripts to keep them focused and accurate. Currently, irrelevant audio recorded after meetings end (e.g., background music) or off-topic discussion (e.g., initial catchup) gets transcribed and included in the AI summary, diluting the meeting's key takeaways.

Desired Behavior

• Ability to select and delete specific segments or lines from a transcript
• Changes to the transcript should be reflected in the AI-generated summary (summary should update based on cleaned transcript)
• Maintain transcript integrity for the core meeting content

Example Scenario

A user ends a Google Meet but Hyprnote continues recording background music, which gets transcribed. The user wants to remove this music transcript segment so it doesn't affect the meeting summary's accuracy.

Similarly, a user might want to remove initial casual conversation ("catchup") that isn't relevant to the meeting's focus.

Benefits

• More accurate and focused transcripts
• Better AI summaries that reflect only the important discussion
• User control over what counts as "meeting content"

[devin-ai-integration]

Codebase Context for Transcript Segment Deletion

Current State

There is already partial infrastructure for editing transcripts, including word-level deletion. Here's what exists and what's missing:

Existing Editing Infrastructure

Edit mode with undo/redo (apps/desktop/src/session/components/note-input/transcript/editing/):

• use-transcript-editing.ts — manages edit mode lifecycle with TinyBase checkpoints for undo/redo
• controls.tsx — renders Edit/Save/Cancel/Undo/Redo buttons
• Edit mode is gated: editable is only true when sessionMode === "inactive" and operations are available (state.ts:68-69)

Word-level mutations (apps/desktop/src/session/components/note-input/transcript/mutations.ts):

• handleDeleteWord(wordId) — deletes a single word from the transcript's words JSON and removes associated speaker hints, then creates a checkpoint
• handleAssignSpeaker(wordIds, humanId) — assigns speaker identity to words
• These are exposed via the Operations type (packages/transcript/src/shared.ts:192-195):

  export type Operations = {
    onDeleteWord?: (wordId: string) => void;
    onAssignSpeaker?: (wordIds: string[], humanId: string) => void;
  };

Right-click context menu (apps/desktop/src/session/components/note-input/transcript/renderer/word-span.tsx:20-31):

• Individual words have a "Delete" context menu item that calls operations.onDeleteWord
• This only works in edit mode (when operations is defined)

Selection menu (apps/desktop/src/session/components/note-input/transcript/renderer/selection-menu.tsx):

• Shows "Copy", "Remove", and "Replace" buttons when text is selected
• "Remove" and "Replace" are only visible when editable is true
• However, the handleSelectionAction in renderer/index.tsx:90-94 only implements "copy" — the "remove" and "replace" actions are not yet wired up:

  const handleSelectionAction = (action: string, selectedText: string) => {
    if (action === "copy") {
      void navigator.clipboard.writeText(selectedText);
    }
  };

What's Missing for This Feature

1. Selection-based removal: The "Remove" button in the selection menu fires onAction("remove", selectedText) but handleSelectionAction doesn't handle it. To implement segment deletion, this handler needs to:

   • Identify which word IDs are covered by the text selection (the DOM selection spans data-word-id attributed spans)
   • Call operations.onDeleteWord for each word, or add a new bulk onDeleteWords(wordIds: string[]) operation to avoid creating a checkpoint per word

2. No segment-level (time-range) deletion: The current Operations type only supports single-word deletion. There's no concept of "delete all words between timestamp X and Y" or "delete this entire speaker segment." The Segment type (packages/transcript/src/shared.ts:32-35) groups words by speaker, so a segment-level delete would need to collect all word IDs in that segment.

3. Replace action: Similarly unimplemented — the selection menu's "Replace" mode captures replacement text but nothing processes it.

AI Summary Pipeline

The AI enhancement pipeline reads transcript words directly from TinyBase at summary generation time:

• enhance-transform.ts:336-376 (collectWordsForTranscripts) — parses the words JSON column from the transcripts table, builds segments, and formats them as Transcript[] for the LLM prompt
• enhance-workflow.ts:83-130 (getUserPrompt) — the transcript content flows into the content field of the prompt context

Key insight: Since the enhancement reads words from TinyBase at generation time, any words deleted via mutations.ts:handleDeleteWord will already be excluded from future AI summaries. No separate change is needed in the AI pipeline — the summary will naturally reflect the cleaned transcript when re-generated.

Recent Related Changes

• 0c744383a (HEAD) — "chore: enable selection from transcript screen": Just merged — makes text selection work on the transcript view even when not in edit mode, adds "Copy" to the selection menu, and gates "Remove"/"Replace" behind editable. This is direct groundwork for this feature.
• f6111c74c — "feat: add transcript copy action (feat: add transcript copy action #4646)": Added the copy action infrastructure.
• 95e1fb80a — "various fixes and cleanup in transcript tab": First version of the selection menu.

Relevant Files

Area	Files
Transcript data types (Rust)	crates/transcript/src/types.rs
Transcript data types (TS)	packages/transcript/src/shared.ts
Selection menu UI	apps/desktop/src/session/components/note-input/transcript/renderer/selection-menu.tsx
Selection action handler	apps/desktop/src/session/components/note-input/transcript/renderer/index.tsx (L90-94)
Word-level mutations	apps/desktop/src/session/components/note-input/transcript/mutations.ts
Edit mode controls	apps/desktop/src/session/components/note-input/transcript/editing/
Transcript screen state	apps/desktop/src/session/components/note-input/transcript/state.ts
AI enhancement transform	apps/desktop/src/store/zustand/ai-task/task-configs/enhance-transform.ts
AI enhancement workflow	apps/desktop/src/store/zustand/ai-task/task-configs/enhance-workflow.ts
Word context menu (delete)	apps/desktop/src/session/components/note-input/transcript/renderer/word-span.tsx
TinyBase persistence (save)	apps/desktop/src/store/tinybase/persister/session/save/transcript.ts
TinyBase persistence (load)	apps/desktop/src/store/tinybase/persister/session/load/transcript.ts

Implementation Considerations

1. Bulk word deletion — Add onDeleteWords(wordIds: string[]) to Operations to handle removing multiple selected words in a single checkpoint, avoiding N separate undo steps.
2. Selection → word ID mapping — The selection menu receives raw text; it needs to resolve which DOM word spans are selected. Word spans render with data-word-id attributes that can be extracted from the selection range.
3. Segment-level delete UX — Consider adding a "Delete segment" action on the segment header (next to the speaker assignment dropdown) for removing entire speaker segments at once.
4. Summary re-generation — After deletion, the user would need to manually re-trigger enhancement ("Enhance" button) to get an updated summary. Consider prompting or auto-triggering re-enhancement after transcript edits.
5. Replace functionality — The selection menu already has a "Replace" mode UI. This would need a new onReplaceWords operation that updates word text in TinyBase.