Three-Tier Correction Engine
Every transcription runs through three matching stages. Corrections only apply at word boundaries, so you won't get mangled partial replacements.
Exact & Phrase Lookup
HashMap lookup for single words and multi-word phrases. Handles known misspellings and abbreviation expansions in O(1) time.
k8s→KubernetesSymSpell Fuzzy Matching
Edit-distance matching with prefix indexing (configurable 0-3 distance). A spell validator gate keeps valid English words from being "corrected."
tenserflow→TensorFlowPhonetic Matching
Catches homophones that edit distance misses. When you say "their" but mean "there," this tier handles it.
their/there/they're→Correct form based on contextPrompt Words
Unlike dictionary corrections, prompt words influence recognition before text is produced. The engine expects these terms and is more likely to hear them correctly.
Whisper treats prompt words as prior context. NVIDIA compiles them into CTC vocabulary models that boost acoustic decoder probability. Same result: better recognition for your specific vocabulary.
Use Cases
- Brand names and product names (e.g., "Kubernetes", "PostgreSQL", "Next.js")
- Personal names and team member names
- Medical or legal terminology
- Company-specific acronyms and jargon
- Foreign words used frequently in your workflow
Dictionary Management
Categories
Group corrections by domain: medical, legal, programming, names. Makes large dictionaries manageable.
Usage Tracking
See which corrections fire most often. Helps you spot what matters and prune what doesn't.
Import & Export
JSON import/export for sharing word lists across devices or with teammates.
Dictionary Packs
Pre-built correction databases you can toggle on or off. Auto-updated when new versions ship.
Frequently Asked Questions
What are prompt words?
Prompt words bias transcription toward specific vocabulary — proper nouns, technical terms, brand names. For Whisper, they're passed as 'previous context'. For NVIDIA, they're compiled into CTC vocabulary models that boost probability at the acoustic decoder level.
Will the dictionary change correctly spelled words?
No. The spell validator gate ensures that SymSpell fuzzy matches are validated against a spell checker — valid English words are not 'corrected' by fuzzy matching. Only genuinely misspelled or misheard words are fixed.
Can I see what was corrected?
Yes. In the live preview, corrected words are shown with gradient color highlighting. Click any highlighted word to see the original transcription before correction.
How do prompt words differ from dictionary entries?
Dictionary entries fix text after transcription (post-processing). Prompt words influence the transcription engine itself, biasing it to recognize specific vocabulary during the acoustic decoding phase.