Feature: Confidence-based highlighting in transcript editor

[erklu]

Context

As part of the transcript editor (see parent issue #3 ), the system should support visual indicators for uncertain transcription segments based on model confidence scores.

With the transition to HuggingFace/PyTorch-based inference, word- or token-level confidence signals can be made available and used to guide user review.

Problem

Users currently have no guidance on where transcription errors are most likely to occur. This leads to inefficient manual review, especially for long recordings.

Proposal

Introduce optional confidence-based highlighting in the editor to visually indicate uncertain words or segments.

Scope

1. Word-level highlighting

• Words below a configurable confidence threshold are visually marked
• Highlighting should be subtle (e.g. underline or soft background), not disruptive

2. Adjustable threshold

• Provide a UI control (e.g. slider or dropdown) to adjust sensitivity
• Users can choose how aggressive the highlighting should be

3. Toggle on/off

• Users must be able to disable confidence highlighting entirely

4. Block-level indicators (optional but recommended)

• Paragraphs/segments with low-confidence content show a summary indicator (e.g. icon or count)

5. Non-destructive

• Confidence data must not alter transcript text
• Highlighting is purely a visual aid

UX considerations

• Avoid red as default color (should not imply "error")
• Prefer neutral or warning tones
• Consider tooltip or side panel for detailed confidence info
• Must not clutter the editor when many words are flagged

Technical considerations

• Backend must provide confidence scores at token or word level
• Mapping between tokens and rendered words must be preserved
• Thresholding should be handled client-side where possible
• Should be compatible with diarization and segmentation model

Out of scope (initially)

• Automatic correction based on confidence
• Confidence-aware re-transcription
• Export of confidence data

Acceptance criteria

• Words below threshold are highlighted in editor
• User can adjust threshold dynamically
• User can toggle feature on/off
• No impact on saved transcript content
• Performance impact is minimal