Skip to content
| Marketplace
Sign in
Visual Studio Code>Other>DictateNew to Visual Studio Code? Get it now.
Dictate

Dictate

Project Anomaly

|
39 installs
| (0) | Free
Offline speech-to-text and paste at the focused text cursor. Included speech model and live wave.
Installation
Launch VS Code Quick Open (Ctrl+P), paste the following command, and press enter.
Copied to clipboard
More Info

Shared indicator notices — 0.18.0

Saved vocabulary and dictation errors use the floating indicator even when VS Code is focused. Notices expand to fit their message and action, replace one another in place, and shrink away when dismissed. Hover to keep a notice readable, for at most 30 seconds. Recording takes priority; an old Undo cannot act on a newer correction. If the native notice cannot be displayed, VS Code shows its normal notification instead. Successful insertion, copying and cancellation remain quiet. Your chosen animation and colours are preserved.

Accurate recording status — 0.14.1

“Recognizing speech” appears only while Dictate is converting a recording. Formatting, insertion and choosing a retry result have their own labels. Finished or unrelated requests cannot leave an old recognition message in the status bar. Click a busy status or press Escape to cancel; successful completion and failures return it to Start. Startup model preparation stays local with the microphone closed.

Faster first recording — 0.14.0

Dictate loads and warms up its included CPU speech model automatically when the extension starts. It uses memory earlier, without opening your microphone or recording audio. Once preparation finishes, the first recording uses the same ready model as later recordings. If you start immediately while it is still preparing, that recording waits for preparation. Light and Verbatim use local rules and need no separate model.

The default keeps the model ready. If you choose to release it after an idle period, the next recording needs to load it again. Normal cancellation prepares a fresh worker; failures remain bounded and report through the next explicit action. No account, network request or download is added.

Recording animations — 0.13.0

iOS 9 wave with Bankinter colours is the default. Run Dictate: Recording Animation and Live Preview to choose Glass island, Classic Siri wave, iOS 9 wave, Listening orb, or any retained animation. The five main controls are Animation, Colours, Size, Movement and Position, with an immediate local preview and optional Advanced controls. The two requested iOS-9 libraries share one waveform choice.

The new animations use the Bankinter orange #FF8D30, yellow #FED227, pale blue #B7F0F0, white #FFFFFF and black #212425. Explicit custom colours can override the theme; choose Colours → Bankinter to restore it. Hover symbols, quiet successful completion, local CPU speech, password/destination protections and native Undo remain. Upstream MIT/BSD notices are included; the Listening orb is original artwork inspired by independently rotating layers.

Wave appearance — 0.10.1

The wave is centered horizontally and vertically in a small capsule. Routine recording shows no brand or status words. Pause/resume, stop and cancel appear as symbols only while hovering, over a dimmed wave; they reserve no extra row or empty space. Loud input uses nearly the full height and accelerates movement. At 200% display scaling, the default is 300 × 55 pixels. Successful insertion/copy closes the indicator without reopening a confirmation. Errors and actionable correction Undo remain available.

Run Dictate: Wave Appearance and Live Preview for the synthetic preview and appearance controls. Choose Aurora, Minimal, Ribbon, Volume History or Sound Bands; high-contrast, static, symbol-only and touch presets are also available. Save sets defaults; Use once affects the next recording. Appearance import/export uses local JSON files and excludes monitor identities/positions and personal data.

The floating wave now uses elapsed-time motion, smooth state/caption transitions, separate Pause/Formatting/result symbols, clipping and unavailable-signal accents, configurable visual sensitivity and cached drawing. Visual sensitivity does not change recorded audio. The header grip moves the wave; Stop/Cancel/Pause controls have separate larger hit targets. Ctrl+wheel resizes over the wave. Dictate: Move, Resize or Hide Wave exposes keyboard-accessible presentation actions; hiding the picture keeps recording active. Error Details is read-only and non-activating. Undo targeting is bounded at 30 seconds.

The preview is an isolated mode of the bundled worker. It opens no microphone, loads no speech model and makes no network requests. It pauses during dictation and closes with its settings panel. Windows motion/high-contrast preferences and real display scaling are respected. Existing target/password/Undo/no-submission protections and local CPU Parakeet/Light/Verbatim remain.

Dictionary transfer — 0.6.9

Use Dictate: Edit Dictionary, then Import dictionary or Export dictionary. The same actions are in the Command Palette. Save edits first. In desktop Dictate, open Dictionary and use Export dictionary / Import dictionary. Transfer the JSON file between computers and import it in the other edition. Only saved words and corrections travel; existing conflicting rules stay. Imports apply to the next recording. Portable files support 1,000 words, 500 corrections and phrases of up to 60 characters; larger dictionaries are rejected without truncation. Export is an explicit local file action; there is no automatic synchronization.

Dictate

Offline speech-to-text for desktop VS Code on Windows x64. The multilingual CPU Parakeet TDT 0.6B v3 speech model is included. No account, speech download, telemetry or network request is needed.

Light cleanup repairs common missing commas and clearly marked sentence breaks in Spanish and English. For example, “He cambiado el modelo pero no funciona” becomes “He cambiado el modelo, pero no funciona.” It uses local rules without another model. Ambiguous breaks cannot always be inferred from text alone; Verbatim leaves recognized punctuation unchanged.

Use

  1. Click the text input or select the Excel cell where you want the result.
  2. Press Ctrl+D to start, wait until the indicator says Recording, then speak. Press Ctrl+D again to stop.
  3. Dictate converts the speech locally, puts the result on the Windows clipboard and sends ordinary Ctrl+V at the current keyboard focus.

The text cursor or selected cell is the destination, not the mouse pointer. You can move the cursor while recording or recognition runs; focus the intended destination before text is ready. Paste replaces selected text as the receiving app normally does. Dictate sends no Enter, Tab or Send command. Excel handles cell paste normally, including its own formula interpretation and Undo. An app that does not accept Ctrl+V will not receive text; the result remains on the clipboard for manual paste.

The same paste path is used inside VS Code editors/chat and outside VS Code. It does not require an application-specific text or selection interface. Copilot needs no accessibility-setting changes. Dictate: Chat Dictation: Start / Stop is a compatibility alias for the same operation.

Ctrl+D uses Windows RegisterHotKey inside and outside VS Code while native shortcuts are enabled, taking priority over editor/browser bindings. Change Dictate: Desktop Shortcut for every input; disable Dictate: Desktop Apps to release it and use VS Code commands/keybindings. There is no global keyboard hook.

Escape cancels. Recordings are limited to 120 seconds. The floating voice wave stays near the bottom of the current monitor, at 320 × 80 pixels on smaller monitors or 480 × 120 on 4K-sized monitors, and never takes focus. The status-bar indicator remains available. Disabling Dictate: Desktop Apps disables outside shortcuts while retaining the wave inside VS Code.

Paste behavior and privacy

The dictated text replaces the previous clipboard contents and remains available there. Windows clipboard history or synchronization, if separately enabled, and other applications may access that clipboard text. Dictate does not create transcript history or enable clipboard history/sync.

Input contents and selections are not read. Windows focus/window identity is checked during the brief paste operation, so a focus change at that moment refuses the shortcut and leaves the text on the clipboard. Missing accessibility/text interfaces do not block paste. Known native/accessibility password fields are rejected; inaccessible custom controls cannot be reliably identified as password fields. Avoid dictating secrets into such controls.

Undo follows the receiving application's normal paste behavior. Dictate verifies exact insertion in supported readable fields up to 600 characters. Unavailable or changed fields remain unverified; a reliably unchanged native field offers Copy recovery. No automatic second paste is sent. Verification briefly reads only that same field, even when learning is disabled. Elevated/security windows may block keyboard delivery. The mouse pointer is never used as an insertion location.

Speech uses included CPU Parakeet TDT 0.6B v3 recognition, automatic spoken-language recognition for Spanish and English, and local Light or Verbatim cleanup. Light leaves one trailing space after completed sentences; Verbatim retains the recognized text except command/control characters, which are removed or replaced with spaces before paste. Line breaks and tabs never become command keys.

The microphone opens only for an explicit recording. Windows shared microphone access allows other apps to use the input too; Dictate does not request exclusive control. Recording ends before recognition starts, and the audio library releases its drivers after stopping or closing the microphone picker. The speech model stays cached for repeat dictation. Audio and recognized text are held temporarily in local worker memory; no audio files or document contents are saved. Preferences use VS Code settings, whose separately enabled Settings Sync may synchronize them. Dictionary words/corrections stay in local extension storage, which Dictate does not register for Settings Sync. The worker blocks sockets, DNS, network paths and external launches. VS Code and receiving apps retain their independent behavior.

Installation and settings

Install the Windows x64 VSIX through Extensions → … → Install from VSIX, then use Developer: Reload Window to activate the installed version. Installation alone can leave an earlier worker running until reload.

Use the Command Palette for Dictate: Choose Microphone and Dictate: Dictation Settings. Speech model, language, cleanup, personal words, outside-app shortcut and wave remain local. Browser VS Code and non-Windows platforms are unsupported.

Dictionary and corrections

Run Dictate: Edit Dictionary from the Command Palette to edit words and corrections. The initial list includes 95 work terms from the server's published Bankinter preset and existing personal words. Version 0.5.7 restores an omitted work term once in previously saved dictionaries while preserving custom words, corrections and older deletions. Private server dictionaries are not automatically copied. Save changes to use them from the next recording; deleted entries stay deleted. Add up to 1,000 words and 500 corrections, with one word/phrase per line and corrections written as bank inter => Bankinter. Saving a correction also adds its written word to the recognition hints. All ordinary word/phrase entries are supplied as local recognition hints. Entries containing slash, colon or a backslash remain saved but are omitted from recognition hints because the decoder reserves those characters; exact corrections still support them. Hints guide recognition; they do not guarantee every spoken term is spelled correctly.

In Light mode, eligible spelling edits to confirmed dictations teach application-specific vocabulary locally. Word changes are checked separately from punctuation, so correcting a name and its sentence ending together can still be learned. New automatic learning does not create replacement rules; use Dictate: Save Correction when you deliberately want one phrase always replaced. Existing rules keep working. Only the proven inserted passage is considered, with a 240-character changed-sentence bound and at most 90 seconds of observation. Supported accessibility text ranges can locate a short passage inside longer chat inputs; native edit boxes retain a 600-character whole-field limit. Passwords, read-only fields, ambiguous insertion, unrelated edits, numbers, common verb-form changes, code and URLs remain excluded. Focus or field changes end observation. Missing, unbounded or slow text access skips learning without blocking ordinary paste. Saved vocabulary notices offer Undo.

Saved rules replace whole phrases once in Light mode. Verbatim skips replacement rules and automatic learning. No editor-wide inspection, clipboard reading or keyboard logging is used. Readable external fields are checked briefly only for the last dictation. Dictionary edits and saved corrections stay local and survive extension updates. Existing Dictate: Words entries are included initially; after the first dictionary save, the new saved list is authoritative.

Validation limits

Version 0.6.4 includes Parakeet TDT 0.6B v3 for faster and more accurate local Spanish/English recognition. The model automatically chooses the spoken language; the language setting is a preference for labeling short ambiguous results, not translation or forced-language decoding. Light/Verbatim, editable work words, saved corrections, automatic editor learning and microphone sharing remain available. Clear Portuguese output is still checked locally, retried once without dictionary hints and blocked if it persists. No model download occurs during use.

These are immediate sample improvements, not a guarantee for every accent, work term or microphone. Very quiet phrases retain the existing bounded volume preparation.

Quiet-speech adjustment runs before the minimum-level check. If no text is returned, the message distinguishes a recording shorter than a quarter-second, insufficient microphone input and audio the model could not recognize. No microphone diagnostics or audio are saved.

Recording locally appears only after the microphone delivers its first 50 milliseconds of fresh samples. If no samples arrive within two seconds, Dictate closes the input and reports that problem. Each recording clears the previous readiness state.

Version 0.6.2 includes the editable correction bank inter => Bankinter in Light mode. Existing dictionaries receive it once if they retain Bankinter and have no custom rule for that phrase. Removed words, custom rules and later correction deletions are preserved; full correction lists are left intact. Verbatim does not apply replacement rules. Other work terms remain recognition hints; ambiguous name replacements are not imposed.

Ctrl+D now starts/stops through one native Windows shortcut inside and outside VS Code, including embedded browser comments. Enable Dictate: Desktop Apps (on by default); Dictate: Desktop Shortcut changes it everywhere. This takes priority over other uses of that shortcut. Escape cancels. The old separate Browser Shortcut setting is no longer used; saved preferences are not deleted. Disabling native shortcuts releases the key and leaves VS Code commands/keybindings available, but browser forwarding is then required. No global keyboard hook, ordinary-key recording or automatic submission is added.

Light cleanup also recognizes bounded English conditional questions, such as "If I change this, does it change that?", and separates a clear opening observation from a following question. It restores punctuation without guessing replacements for misheard ordinary words.

Dictionary corrections protect negative contractions: a name rule such as Don => Dawn leaves don't alone, while still correcting possessives. An explicit full-contraction rule still applies. Automatic editor learning skips changes between common English/Spanish verb forms, such as made and make; use Save Correction for a deliberately chosen replacement.

Saved automatic and manual corrections show their exact spelling change in a VS Code notification with Undo. A non-activating desktop notice also shows Dictionary saved and Undo for 15 seconds, without opening VS Code or stealing typing focus. Undo removes only that action’s changes and restores an overwritten manual rule, preserving other dictionary entries.

Dictionary and retry improvements (0.7.0)

Run Dictate: Edit Dictionary to search words, application rules and source labels. Use excel.exe | recognized phrase => written phrase for an application-specific correction; omit the prefix to apply it everywhere. Application rules take priority over global rules. Light cleanup applies corrections and spoken formatting; Verbatim keeps recognized words. Say “new paragraph” / “nuevo párrafo” or “bullet point” / “punto de lista” to format text. Quoted text, code and URLs remain protected.

Dictate: Undo Last Dictionary Import reverses unchanged additions from the last import in this session, preserving later edits. Source labels distinguish manual, learned, imported and bundled corrections; older entries without records show unknown. Export/import transfers these labels and application rules with words.

Dictate: Retry Last Recording (Copy Result) reprocesses the recording with the current dictionary and corrections. The latest recording stays only in memory for five minutes; new recording, cancellation and worker exit discard it. Retry copies the result, keeping previously pasted text. It uses the same fast decoder as ordinary dictation.

Enable Dictate: Share Local Dictionary and the matching desktop option in Dictionary → Advanced to share local entries on this Windows account. Sharing is off by default. Existing shared corrections win on first connection; later changes are exchanged on use, without a background watcher. No account or network is needed.

Text recovery and explicit entry commands (0.8.0)

Run Dictate: Copy Last Recognized Text after a delivery failure. The last recognized result remains only in memory for five minutes; a new recording, explicit Cancel, expiry or shutdown clears it. This action only copies, never pastes or submits. Audio Retry remains separate.

Light supports bounded digits … end digits / dígitos … fin dígitos, spell … end spell / deletrea … fin deletreo, and snake case, camel case or pascal case … end identifier / fin identificador commands. For example, digits one five point zero end digits writes 15.0; camel case account balance end identifier writes accountBalance. Unknown commands, quoted text and existing technical literals stay literal; Verbatim skips conversion. Dates and amounts are not guessed. Application-specific correction rules are chosen again for the receiving app from original recognized words; an app change during the brief delivery operation blocks paste and preserves recoverable text.

Choose Dictate: Entry Mode in VS Code Settings for whole-recording digits, spelling, snake_case, camelCase or PascalCase entry with Light cleanup. natural is the default. Unsupported wording stays unchanged; Verbatim ignores the setting. Spoken opening/closing commands can be misrecognized, so the explicit setting avoids depending on a closing word.

Version 0.8.1 also preserves completed recognition after an automatic worker/helper exit or a processing-wave failure. Explicit Cancel still clears recovery; no automatic retry, paste or submission occurs.

Special commands and recovery

Run Dictate: Dictation Settings for the Special commands section and English/Spanish examples. Choose Light cleanup and a mode. Money + es converts “mil doscientos euros con cincuenta céntimos” to 1.200,50 €; en produces €1,200.50. Say the amount itself. Use once changes only the next recording. Dates require a named month; unknown or ambiguous phrases are left unchanged.

The command palette also offers Pause / Resume, Switch to Previous Microphone, Copy Original Recognized Words, Reformat Last Result and Paste Last Result. Results and retry audio expire after five minutes. Reformat copies text using current settings without another recognition pass. Paste Last Result is explicit and checks the current destination; it never sends a message. Retry offers a comparison before copying.

The dictionary supports row editing, disabling, a sentence tester and selected imports with conflict preview. Project vocabulary uses a group you explicitly name in a trusted workspace; no files are scanned. Application and language rules apply only in their scope. Project/language and disabled rules stay local during sharing/export. Automatic correction notices offer Undo and never learn; the wave's Undo is complemented by the Never Learn command.

Advanced settings include 30/60/120-second recording limits, 1/2/4 CPU threads, an optional ready sound and optional model release after 5/10/20 idle minutes. Idle release and ready sound are off by default. Paused time does not count toward the capture limit. A disconnected input is retried once using the same saved device identity; another microphone is never silently selected.

Spanish primary controls follow the chosen interface language. VS Code translates command titles and configuration descriptions using its own display language. System/driver errors may retain their original wording.

Favourite vocabulary

Open Dictate: Dictionary and use ☆ Favourite beside a word; ★ Unstar removes its priority. Stars put applicable words first in recognition hints without changing replacement rules or model scores. Application/project/language scopes still apply. Stars stay local and are retained by complete local dictionary exports.

  • Contact us
  • Jobs
  • Privacy
  • Manage cookies
  • Terms of use
  • Trademarks
  • Your Privacy Choices
  • Consumer Health Privacy
© 2026 Microsoft