Feature Request: AI Review Assistant for Context-Aware Transcript Correction Suggestions #74
Description
Background
Completed transcripts often contain domain-specific terminology errors, inconsistent naming, or likely mishearings that are difficult to detect with simple post-processing rules.
A review assistant could help users improve transcript quality by analyzing the completed transcript in context and suggesting possible corrections, while leaving all final decisions to the user.
This would be especially valuable for research material, teaching content, accessibility work, and transcripts containing specialized terminology.
Requested functionality
Add a Review Assistant that can be launched manually from a completed transcript.
The assistant should:
- analyze the completed transcript before review starts
- use the transcript language and communicate in that same language
- propose a likely domain based on the transcript content
- allow the user to confirm or change the proposed domain before suggestions are generated
- review the transcript in context and suggest possible corrections one by one
- never apply changes automatically
Expected behavior
Step 1: Start review manually
The user opens a completed transcript and starts the assistant explicitly, for example through a button such as:
- Start Review Assistant
This should be a deliberate review action, not an automatic post-processing step.
Step 2: Domain suggestion
The assistant analyzes the transcript and presents a suggested domain, for example:
- Medicine
- Law
- Humanities
- Academic administration
- Technology
- General academic
The domain should be presented as a suggestion, not as a fixed classification.
Example UI actions:
- Continue
- Change domain
Step 3: User confirms or changes domain
Before any correction suggestions are shown, the user can either:
- continue with the suggested domain
- select a different domain manually
This is important because automatic domain detection may be imperfect, and the user may know the material better than the system.
Step 4: Suggest corrections one by one
After domain confirmation, the assistant reviews the transcript and presents possible corrections individually.
Each suggestion should include:
- original text
- suggested replacement
- short rationale
Example UI actions:
- Accept
- Dismiss
- Skip
No suggestion should be applied unless the user explicitly accepts it.
Types of suggestions
The first version should focus on conservative, review-friendly suggestions such as:
- terminology correction
- inconsistent spelling of names or terms
- likely proper-name correction
- suspicious word choice based on context
- inconsistent usage of the same term across the transcript
The assistant should behave as a review tool, not as an automatic rewriting tool.
Important constraints
The assistant must not:
- rewrite the whole transcript
- silently apply edits
- aggressively normalize spoken language into formal prose
- act as an automatic editor
This feature should support transcript review, not replace it.
Language behavior
The assistant’s explanatory text should use the same language as the transcript.
Examples:
- Swedish transcript → assistant communicates in Swedish
- English transcript → assistant communicates in English
At the same time, all buttons, actions, and system UI labels should remain in English for consistency with the rest of the product.
Why this matters
This feature would provide a safer and more trustworthy first step toward smarter transcript improvement than automatic dictionary replacement.
Benefits include:
- better support for domain-specific terminology
- improved consistency across long transcripts
- reduced risk of silent or incorrect edits
- stronger user trust through explicit approval of every accepted change
- better fit for research and other sensitive material
Out of scope for first version
The first version does not need to include:
- automatic acceptance of suggestions
- bulk accept actions
- custom domain creation
- free-form rewriting of transcript sections
- organization dictionaries
- user terminology settings
These can be considered in later iterations.
Acceptance criteria
- User can manually start the Review Assistant from a completed transcript
- Assistant proposes a likely domain before starting review
- User can change the proposed domain
- Assistant communicates in the same language as the transcript
- Buttons and UI controls remain in English
- Suggestions are presented individually
- Each suggestion can be accepted, dismissed, or skipped
- No changes are applied automatically
- The assistant does not rewrite the full transcript