Tiger Transcribe — Offline Dictation

Dictation Pad

Idle. Click Start Dictation to begin.

Quick Voice Commands

new paragraphline break
new linesingle break
period / commapunctuation
question markpunctuation
exclamation pointpunctuation
delete last wordundo
delete thatclear last utterance
insert datetoday's date
copy to clipboardaction
scratch allclear pad
select last word / sentence / paragraphhighlight
select last N wordse.g. "three"
select all / select thathighlight
copy / cut / deleteact on selection
acceptdeselect

How it works

Every keystroke of your voice is processed locally on this device using the Vosk speech engine compiled to WebAssembly. No audio ever leaves the browser. The model is cached once, then works entirely offline.

Custom Dictionary

Add proper nouns, jargon, or acronyms the recognizer might miss. Words added here are injected directly into the recognizer's grammar, biasing it toward your vocabulary.

Import JSON

Note: custom vocabulary is applied when the recognizer loads. After adding words, click Reload Recognizer or restart dictation for changes to take effect.

Macros & Voice Commands

Bind spoken phrases to text expansions or actions. When the recognizer hears your phrase, it fires instead of transcribing it literally.

About this application

Tiger Transcribe is a fully offline dictation tool. After the model loads once, the page can be saved to disk and opened from file:// — everything runs entirely in your browser.

Technical profile

Engine — Vosk (Kaldi) compiled to WebAssembly via vosk-browser. Runs entirely in your browser — no audio ever leaves your device.
Model — vosk-model-en-us-0.22-lgraph (≈ 128 MB, cached in IndexedDB after first load).
Audio — Web Audio API + AudioWorklet, resampled to 16 kHz mono.
Storage — IndexedDB for custom dictionary, macros, and transcript autosave.
Privacy — Fully offline after first load. No audio leaves your device. No analytics. No telemetry.

Keyboard shortcuts

⌘/Ctrl+Shift+M — Toggle mic (works anywhere) · Esc — Stop dictation · Space (pad unfocused) — Start/Stop · ⌘/Ctrl+Shift+C — Copy all · ;; — Insert today's date

First-run note

The first time you start dictation, the app will download the speech model (~128 MB) and cache it. Subsequent runs work offline. To make fully portable, save this page (Ctrl+S) after the model has been cached.

Reset & troubleshooting

If recognition is misbehaving or you want to start fresh, you can wipe all local data — custom dictionary, macros, saved transcript, and the cached ~128 MB speech model. The page will reload afterward and the model will re-download on next dictation.

Note on file://: some browsers restrict Web Workers and blob URLs when this page is opened directly from disk, which can break the recognizer. If you see URL or security errors in the console, serve the file over a local HTTP server instead (e.g. python3 -m http.server).