Speaker Correction

HeyDonna provides tools to fix speaker labels in your transcript — swap mislabeled speakers, replace names, and let AI automatically correct attribution errors during proofreading.

Swap Speakers

When two speakers are mislabeled (e.g., the attorney's lines are tagged as the witness and vice versa), you can swap them instantly.

1. Select the text containing the mislabeled speakers
2. Press Cmd+Shift+S (Windows: Ctrl+Shift+S)
3. If two speakers are found, they swap automatically
4. If three or more speakers are in the selection, a dialog appears so you can pick which pair to swap

A confirmation appears showing which speakers were swapped (e.g., "Swapped MR. SMITH and THE WITNESS").

Replace Speaker

When a speaker label needs to change to a different name entirely:

1. Select text containing the speaker you want to replace
2. Press Cmd+Alt+R (Windows: Ctrl+Alt+R)
3. Choose the replacement speaker from the dialog
4. All instances of the selected speaker within your selection are updated

This is useful when a speaker was initially labeled as "S1" and you want to replace it with their actual name.

Automatic Speaker Correction

During AI Proofreading, Donna automatically analyzes speaker attribution using conversation context:

• Q&A alternation — In depositions, questions come from attorneys and answers come from the witness. Donna checks this pattern and fixes violations.
• Role consistency — Judges give rulings, attorneys ask questions, witnesses answer. Donna ensures labels match these roles.
• Short interjections — Brief responses like "Yes," "Correct," or "Objection" are attributed to the correct speaker based on context.
• Speaker drift — When labels appear to have swapped partway through, Donna corrects from the swap point forward.

Managing Speaker Names

When you create a project, you can assign real names to speaker labels:

• S1 → MR. SMITH
• S2 → THE WITNESS
• S3 → THE COURT

These mappings are used throughout the editor and during AI processing. The AI uses the speaker names you've assigned to make better attribution decisions.

Suggest Names

Open Speakers in the editor's side panel and click the small outlined Suggest names button centered above the speaker cards (sparkle icon and text label) to ask AI to identify names from the current transcript. Opening or reopening the panel is free; clicking Suggest Names uses credits through speaker correction billing. It never runs automatically.

1. Review the proposed names beside each speaker's current name. Edit a suggestion before accepting it.
2. Select one name, or use Select all to select all eligible names, then click Apply. Roles can be reviewed and edited for legal projects only; general projects show names only.
3. Click Dismiss to discard remaining suggestions. Names you already applied stay in your draft.
4. Click Save Changes to save accepted names and update their labels in the transcript. Receiving suggestions or clicking Apply alone does not save them.

Existing manually assigned names, resolved names, and non-generic labels are preserved. Speaker IDs never change. If the transcript, saved version, or speaker list changes after you request suggestions, Apply shows a message instead of applying outdated names. Dismiss them and click Suggest Names again. If the request fails or no confident names are found, your transcript is unchanged and you can name speakers manually.

Voice samples

Open View → Speakers to listen to participant samples inside the existing Speakers panel. Opening the panel uses no credits and never requests AI suggestions. Formatted Q/A participants are listed even when they have no usable sample. Existing marked speakers keep their name, role, save and suggestion controls; mixed transcripts show both kinds of participants without duplicating rows.

Every participant and speaker label uses the same card layout: the S1 / S2 / S3 badge, person icon, name, and role are centered on one line. Long names truncate with an ellipsis; hover the name to read it in full. Number and role badges remain unclipped. Existing numbered speaker IDs retain their numbers; other identities receive the lowest unused numbers in stable ID order. These numbers are display-only and never change saved speaker identities or sample attribution.

Cards share compact Play, Go to turn, Edit, and Delete icon buttons, in that order, left-aligned below the name with tooltips and toolbar-sized compact buttons. Hover or focus an icon to see its action. Edit and Delete manage transcript speaker labels; clicking either for a formatted-only participant explains this without changing the transcript. Read-only projects likewise explain unavailable management actions on click. Save Changes and Apply are filled primary actions on the right. The Save button shows Save Changes beside its icon and Saving... while saving.

• Play (triangle icon) listens to an aligned sample, up to 15 seconds, from that participant.
• Go to turn pauses and seeks the audio and selects the matching words, keeping the panel open.
• Close the panel to return to the full editor width without changing your transcript.

For formatted depositions, questions use the current uniquely matched BY examiner and answers use an explicitly introduced witness. A new examination clears the examiner until another BY line; a new witness replaces the previous witness. Printed Q/A, speaker labels, and transcript text remain unchanged.

For a participant without any linked speaker label, an exact full-name prefix at the start of a transcript paragraph (case-insensitive) can identify a sample. The first qualifying timed speech span of at least 3 seconds in one paragraph is used, capped at 15 seconds; the printed name is excluded. Short paragraphs are never combined. Linked speaker and Q/A samples retain their existing 10–15-second selection.

No sample status line is shown. Play and Go to turn stay enabled; when speech is missing, clicking shows only the toast “No speech from this speaker in the transcript yet.” Missing or ambiguous names never guess an identity; an appearance alone is not speech. Audio or transcript readiness errors remain distinct. On transcripts with speaker labels, the same panel retains the existing name and role management controls.

Tips

• Swap first, then proofread — If you notice a large section with swapped speakers, use Cmd+Shift+S to fix it before running AI proofreading
• Name your speakers — Assigning real names to S1, S2, etc. in project settings helps the AI make better speaker attribution decisions
• Check depositions carefully — AI speaker correction is designed for legal depositions with Q&A patterns. It may be less accurate for informal conversations or multi-party discussions