VoiceInk is a native macOS application that transcribes what you say to text almost instantly. This fork also contains an in-progress native Windows implementation under VoiceInk.Windows/.
The Windows fork is developed as a free/open-source implementation. It has no licensing gates, trials, purchase prompts, paid feature locks, commercial telemetry, or private paid updater channel.
The project goal is to make privacy-focused voice-to-text software that is efficient, understandable, and source-runnable.
- 🎙️ Accurate Transcription: Local AI models that transcribe your voice to text with 99% accuracy, almost instantly
- 🔒 Privacy First: 100% offline processing ensures your data never leaves your device
- ⚡ Power Mode: Intelligent app detection automatically applies your perfect pre-configured settings based on the app/ URL you're on
- 🧠 Context Aware: Smart AI that understands your screen content and adapts to the context
- 🎯 Global Shortcuts: Configurable keyboard shortcuts for quick recording, with Windows primary/secondary Toggle, Push to Talk, and Hybrid modes
- 📝 Personal Dictionary: Train the AI to understand your unique terminology with custom words, industry terms, and smart text replacements
- 🔄 Smart Modes: Instantly switch between AI-powered modes optimized for different writing styles and contexts
- 🤖 AI Assistant: Built-in voice assistant mode for a quick chatGPT like conversational assistant
Build the macOS app by following BUILDING.md. Build the Windows fork by following the Windows Fork Development section below.
- macOS 14.4 or later
- Building from Source - Detailed instructions for building the project
- Contributing Guidelines - How to contribute to VoiceInk
- Code of Conduct - Our community standards
The Windows implementation lives under VoiceInk.Windows/ and is separate from the macOS SwiftUI app.
- Windows 11
- .NET 10 SDK
- Visual Studio with Windows App SDK support, or equivalent Build Tools
- A local whisper.cpp-compatible
.binmodel file
Run these commands from the repository root, so the VoiceInk.Windows\... paths resolve correctly:
dotnet restore VoiceInk.Windows\VoiceInk.Windows.sln
dotnet build VoiceInk.Windows\VoiceInk.Windows.sln -c Debug -p:Platform=x64Run this command from the repository root:
dotnet run --project VoiceInk.Windows\src\VoiceInk.Windows.App\VoiceInk.Windows.App.csproj -c Debug -p:Platform=x64If dotnet reports NETSDK1045 because the system SDK is older than .NET 10, either install the .NET 10 SDK or run with a local .NET 10 SDK executable. In this worktree, the bundled SDK command is:
& ..\.dotnet-sdk-10\dotnet.exe run --project VoiceInk.Windows\src\VoiceInk.Windows.App\VoiceInk.Windows.App.csproj -c Debug -p:Platform=x64If the command exits without showing a window from a very long checkout path, the Windows App SDK bootstrapper may be hitting a path-length limit. Move the checkout to a shorter path, or map the long parent directory to a temporary drive with subst and run from there:
cmd /c 'subst W: "C:\path\to\worktrees-parent"'
Set-Location W:\your-checkout
& W:\.dotnet-sdk-10\dotnet.exe run --project VoiceInk.Windows\src\VoiceInk.Windows.App\VoiceInk.Windows.App.csproj -c Debug -p:Platform=x64Remove the temporary drive mapping with:
cmd /c "subst W: /D"The Windows fork includes a repo-local packaging script for an unpackaged, self-contained developer ZIP. Run it from the repository root:
.\VoiceInk.Windows\scripts\package-dev-zip.ps1 -DotNetPath "..\.dotnet-sdk-10\dotnet.exe"The ZIP is written to VoiceInk.Windows\artifacts\dev-zip\VoiceInk-Windows-dev-win-x64.zip. Extract it and launch VoiceInk.Windows.App.exe. On first run, configure a whisper.cpp-compatible GGML .bin model such as C:\Models\ggml-base.en.bin; app data remains under %LocalAppData%\VoiceInk.Windows. The dev ZIP includes the repository GPL license text as LICENSE.txt.
This dev ZIP is for source-built testing. It does not create Start Menu shortcuts, register uninstall entries, sign binaries, or install an MSIX/package identity.
For a per-user source-built install from the dev ZIP, print the install plan first:
.\VoiceInk.Windows\scripts\install-dev-zip.ps1 -PackagePath VoiceInk.Windows\artifacts\dev-zip\VoiceInk-Windows-dev-win-x64.zipThen run the bounded current-user install only when you are ready to extract under %LocalAppData% and create a Start Menu shortcut:
.\VoiceInk.Windows\scripts\install-dev-zip.ps1 -PackagePath VoiceInk.Windows\artifacts\dev-zip\VoiceInk-Windows-dev-win-x64.zip -ExecuteThe matching uninstall helper is also plan-first:
.\VoiceInk.Windows\scripts\uninstall-dev-zip.ps1
.\VoiceInk.Windows\scripts\uninstall-dev-zip.ps1 -ExecuteThese dev ZIP install helpers do not create certificates, import certificates, write machine-wide install locations, or install MSIX packages.
The Windows fork also includes the open-source MSIX packaging foundation: VoiceInk.Windows\src\VoiceInk.Windows.App\Package.appxmanifest plus a repo-local packaging script. MSIX packages must be signed. The repo does not create certificates, import certificates, or store signing secrets; provide a maintainer-owned .pfx when building a package:
Before using a real certificate, run the non-mutating preflight from the repository root to verify paths, MSIX publish properties, artifact containment, and the manual smoke command sequence:
.\VoiceInk.Windows\scripts\package-msix.ps1 -Preflight -DotNetPath "..\.dotnet-sdk-10\dotnet.exe"To review the full open-source release packaging checklist without creating certificates, publishing packages, installing packages, or reading signing secrets, run:
.\VoiceInk.Windows\scripts\test-release-readiness.ps1.\VoiceInk.Windows\scripts\package-msix.ps1 -DotNetPath "..\.dotnet-sdk-10\dotnet.exe" -PackageCertificateKeyFile "C:\Path\VoiceInk.Windows.Signing.pfx" -TimestampServerUrl "https://timestamp.acs.microsoft.com" -TimestampDigestAlgorithm SHA256 -ValidateAfterBuildIf the certificate has a password, add -PackageCertificatePassword "<password>" from a secure local shell. -TimestampServerUrl defaults to https://timestamp.acs.microsoft.com and -TimestampDigestAlgorithm defaults to SHA256, matching the timestamped signing guidance for MSIX packages, and both can be overridden by maintainers. -ValidateAfterBuild runs the non-installing MSIX artifact validator after a successful signed publish and fails if the package is missing, ambiguous, or malformed. Artifacts are written under VoiceInk.Windows\artifacts\msix. On a test machine where the certificate is trusted, smoke the package with:
.\VoiceInk.Windows\scripts\test-msix-package.ps1 -PackagePath <path-to-msix>
.\VoiceInk.Windows\scripts\smoke-msix-install.ps1 -PackagePath <path-to-msix>
Add-AppxPackage -Path <path-to-msix>
Get-AppxPackage VoiceInk.Windows
Remove-AppxPackage -Package <package-full-name>smoke-msix-install.ps1 prints the signed install/query/uninstall smoke plan by default. On a disposable test install where the signing certificate is already trusted, pass -Execute to run Add-AppxPackage, verify the package with Get-AppxPackage, and remove that exact installed package with Remove-AppxPackage; before executing, the helper checks Get-AuthenticodeSignature and refuses to install when the signature status is not Valid. package-msix.ps1 -Preflight does not create output folders, delete output folders, publish, sign, create certificates, import certificates, install packages, uninstall packages, or read certificate passwords. package-msix.ps1 -ValidateAfterBuild still requires a maintainer-owned signing certificate and validates the produced artifact without installing or launching it. test-msix-package.ps1 only inspects the artifact under VoiceInk.Windows\artifacts; it does not install, uninstall, launch, sign, create certificates, import certificates, or cryptographically verify the package signature.
test-release-readiness.ps1 only verifies that the release packaging scripts/assets are present and prints the ordered packaging commands; it does not run publish, signing, install, uninstall, launch, or certificate-store operations. The App Installer helpers are also non-mutating: write-appinstaller.ps1 and test-appinstaller.ps1 require the MainPackage URI to point at a .msix or .msixbundle artifact, so a release manifest cannot accidentally reference a ZIP, HTML page, or documentation URL.
The optional WinGet helpers are also non-mutating: write-winget-manifest.ps1 writes package/default-locale/installer YAML under VoiceInk.Windows\artifacts with InstallerType: msix, and test-winget-manifest.ps1 validates the generated metadata without running winget, submitting manifests, downloading installers, installing packages, or changing certificate trust.
write-release-checksums.ps1 creates a deterministic SHA256SUMS.txt file for staged artifacts under VoiceInk.Windows\artifacts using Get-FileHash -Algorithm SHA256; it refuses files outside the artifact root and does not publish, install, sign, or mutate certificate trust.
For release transparency, generate an SPDX 2.2 SBOM after final artifacts are staged and publish it next to the signed package/checksums. The release readiness report includes a non-mutating Microsoft SBOM Tool (sbom-tool) reference; the repo does not download SBOM tools, publish artifacts, install packages, sign packages, create certificates, import certificates, or trust certificates.
For release certification evidence, run the Windows App Certification Kit only on a disposable or prepared Windows runner with an active user session. The release readiness report documents both appcert.exe test -packagefullname for an installed package and appcert.exe test -appxpackagepath for a signed package file. The manual self-hosted installer-smoke workflow has a default-off run_wack input and uploads wack-report.xml with installer evidence when explicitly enabled; no local active-desktop WACK run is performed by repo scripts. After downloading the installer-smoke-evidence artifact from the disposable runner, validate the expected summary and optional WACK XML evidence with .\VoiceInk.Windows\scripts\test-installer-smoke-evidence.ps1 -EvidenceRoot <path-to-installer-smoke-evidence> -RequireInstallSmoke -RequireWackReport.
On first launch, VoiceInk for Windows opens a setup dialog for the local whisper model path, recommended model download, microphone settings/input, primary shortcut, microphone health, and a short try-it flow. The setup dialog also shows a readiness checklist for model path, primary shortcut, microphone device detection, Windows microphone privacy review, and the app microphone capability; its programmatic controls and readiness list regions expose explicit UI Automation names for screen readers. It never changes Windows privacy settings automatically, but it links to ms-settings:privacy-microphone when review is needed. After setup, use the sidebar to move between Dashboard, Transcribe Audio, History, Metrics, AI Models, Enhancement, Power Mode, Permissions, Audio Input, Dictionary, Settings, and About / Open Source. The Permissions section mirrors the macOS readiness page with cards for keyboard shortcut, microphone access, text insertion, and screen context; card actions route to VoiceInk Settings/Enhancement or to Windows microphone privacy settings. In AI Models, local Whisper catalog cards show language, size, speed, accuracy, description, downloaded/default state, direct download progress, set-default, and Show in Explorer actions. The selected model path is checked for .bin extension, file presence, empty files, and suspiciously tiny files before warmup, Set as Default, or Show in Explorer actions proceed. Catalog downloads use the open-source whisper.cpp GGML files from Hugging Face, save .bin files under %LocalAppData%\VoiceInk.Windows\Models, add them to imported local models, and select the downloaded model as the default path. The Transcription Language control follows the selected model: English-only .en models force English, while multilingual and imported local models offer auto-detect plus Whisper language codes. The Prewarm Local Model setting defaults on and schedules a conservative Whisper.net preload for the selected local model on startup, Windows resume, model import/selection, catalog download, and provider settings save; Warm Up Selected Model runs the same preload manually. Show Live Transcript Preview persists the macOS-style preview preference. When Deepgram, AssemblyAI, Soniox, Speechmatics, or Cartesia is selected with a saved key, the recorder can stream copied 16 kHz mono PCM chunks for interim preview text; final insertion/history still come from the stopped-recording transcription result. Windows skips macOS Core ML encoder downloads because Whisper.net uses the GGML .bin model directly. You can still enter a custom local whisper model path, use Import Model to add an existing whisper.cpp .bin file to the imported local models selector, and use Open GGML Model Downloads to open the upstream model page in the default browser. In Enhancement, you can enable the default-off AI enhancement pipeline, choose an AI provider preset, configure an OpenAI-compatible chat-completions endpoint/model, choose or edit a prompt, create local custom prompts, persist trigger words, turn on read-only Clipboard Context, set timeout/short-phrase/retry behavior, and save or clear the selected provider API key in Windows Credential Manager. Predefined prompt text stays source-controlled, but predefined trigger words can be edited locally; custom prompt title, icon, description, instructions, trigger words, and system-instruction wrapping are stored in JSON settings. Provider presets currently include Custom OpenAI-compatible, Cerebras, Groq, Gemini, Anthropic, OpenAI, OpenRouter, Mistral, Ollama, and Local CLI. The hosted presets fill the macOS-aligned endpoint/model defaults and store keys independently per provider. Anthropic uses its native Messages API with x-api-key and anthropic-version headers. Local CLI uses the endpoint field as a Windows command template, passes VOICEINK_SYSTEM_PROMPT, VOICEINK_USER_PROMPT, and VOICEINK_FULL_PROMPT, and uses stdout as the enhanced text. OpenRouter can refresh model IDs from https://openrouter.ai/api/v1/models; Ollama uses the native local endpoint http://localhost:11434/api/chat, sends stream: false chat messages without an API key, and can refresh installed local models from Ollama /api/tags. Remote enhancement endpoints must use HTTPS; plain HTTP is accepted only for localhost development endpoints. VoiceInk rejects enhancement endpoint URLs that contain embedded credentials or common key/token query parameters so API keys stay out of JSON settings. When enhancement runs, VoiceInk also attempts read-only selected text capture through Windows UI Automation and appends available selected text inside <CURRENTLY_SELECTED_TEXT> tags. Clipboard Context reads current plain text from the clipboard only for the enhancement prompt, appends it inside <CLIPBOARD_CONTEXT> tags, skips empty or unavailable clipboard content, and does not modify clipboard contents. Screen OCR Context is default-off and can read local screen text into <SCREEN_OCR_CONTEXT>; use Select Region to drag a visual full-screen picker, or edit the numeric region fields directly, when OCR should be constrained to a specific area. In Power Mode, add enabled rules that match the active Windows process name and/or title bar text, optionally mark one rule as the default fallback, and set session-only overrides for model path, language, enhancement, prompt, trailing-space behavior, filler cleanup, punctuation cleanup, and lowercase output. VoiceInk captures the matching Power Mode at recording start and records the rule name/emoji in History without rewriting base settings. In Dashboard, click Start Recording, speak, then click Stop And Insert. You can also press Ctrl+Alt+Space to toggle recording by default when that global hotkey is available. During recording and processing, Settings Interface > Recorder Style switches the always-on-top recorder between the bottom mini pill and a Windows top-center notch-style pill. Both styles show state text, elapsed time, live microphone level bars while recording, and processing pulse bars otherwise. While recording, compact Stop and Cancel buttons route through the same guarded commands as shortcuts, tray commands, and the main window. The Prompt button opens a no-activate in-window chooser with an AI Enhancement toggle and prompt rows; selecting a prompt enables enhancement and persists the choice. The Power Mode button opens a matching chooser with Auto and enabled rules, and persists the explicit selection. The Prompt and Power chooser panels also open on pointer hover, stay visible while the pointer is over the button or panel, and close after a short delay when the pointer leaves both. Recorder changes made during capture affect the active result because completion reloads current Prompt/Power settings against the original recording target, and the recorder uses a no-activate mouse hook so clicking it does not take focus away from the dictation target. When live preview is enabled and provider partial text is available, the recorder expands to show that text near the controls; the text is in-memory only and clears at recording boundaries. In Transcribe Audio, choose or drag-and-drop one or more supported audio/video files, then start the queue to import them into app-owned WAV recordings, transcribe them with the selected transcription provider, optionally enhance the cleaned text when Enhancement is enabled/configured, apply dictionary/cleanup settings, save completed rows to History, and copy or save the selected completed queue text as TXT or Markdown, or explicitly enhance it with the current enhancement provider without changing the global Enhancement toggle. Pending, failed, and interrupted processing queue items are restored from a local JSON snapshot on restart when their source files still exist. Completed recorder, Transcribe Audio, and retry rows also write local session metrics to metrics.db; the Metrics section shows sessions recorded, words dictated, words per minute, keystrokes saved, estimated time saved, and transcription/enhancement model performance. Metrics can be filtered by Last 7 Days, Last 30 Days, This Year, or All Time, and exported as a local CSV through the save picker. In Settings, shortcut fields are read-only recorder targets with Record buttons; click Record, press a modifier+key combination, and VoiceInk normalizes the captured shortcut to the same format used by global registration. Bare Escape clears the focused shortcut field. Export Settings writes a local JSON backup for general settings, custom prompts, Power Mode rules, imported model references, and dictionary entries; Import Settings lets you choose categories before applying them. Provider API keys stay in Windows Credential Manager and are never written to the backup file, and any credential-bearing provider endpoint is cleared from the backup, so imported provider settings require reconfiguring keys locally. Settings also includes Clipboard controls for keeping the previous clipboard content, choosing the restore delay, and switching between the default clipboard paste method and Windows Direct Text. Clipboard restore uses a VoiceInk paste-session marker and restores only if the clipboard still contains VoiceInk's transient paste payload, so later user clipboard changes are left alone. Settings also includes local-only Privacy controls: transcript cleanup deletes old history rows and their app-owned audio files, audio cleanup deletes old app-owned audio files while preserving transcript text, and audio cleanup is hidden while transcript cleanup is enabled. When enabled, cleanup runs on launch, after completed recordings, and on a daily in-app timer; manual Run ... Now actions remain available. Cleanup only deletes root-level app-created WAV files in VoiceInk's recordings directory. The General group includes Launch at Login, which registers the current source-built executable under the current user's Windows Run key with a --voiceink-startup argument so login launches start hidden to the tray, plus Reset Onboarding, which marks the first-run setup to show again on next launch after confirmation. The Windows shell also exposes configurable key+modifier shortcuts for primary and secondary recording toggle, paste last, paste last enhanced, retry last transcription, cancel recording, open history, and quick add to dictionary. A Windows tray icon stays available while the app is running; its menu can show or hide the shell, start or stop recording, switch model/provider/language/prompt/Power Mode/audio/context settings, open Quick Add to Dictionary, focus History, open key settings pages, or quit VoiceInk. Retry Last reuses the latest completed history row with a saved audio file, applies the current transcription provider, dictionary, and cleanup settings, saves the retried row, records a retry metric, and copies the retried text to the clipboard. Cancel Recording stops the active recorder, keeps the captured audio file, and saves a canceled history row without adding a session metric. Open History restores the main window and focuses the History section. Quick Add opens a small Vocabulary/Word Replacement dialog that mirrors the macOS quick-add panel intent.
AI Models library action rows, storage guidance rows, and local Whisper catalog cards expose explicit UI Automation names so screen readers announce the same state, source, model metadata, speed, accuracy, and guidance shown visually.
VoiceInk registers a single Windows App SDK app instance key at launch. A second normal launch redirects activation to the existing process, restores the hidden/minimized main window, and exits the duplicate process so the tray icon, recorder, and global shortcuts have one owner.
The OpenAI enhancement preset follows the macOS request-temperature rule: gpt-5 model names use 1.0, while other OpenAI models and Custom OpenAI-compatible endpoints use the existing 0.3 default.
The tray icon tooltip includes the current VoiceInk status plus concise taskbar corner overflow pinning guidance, while the tray menu keeps the longer visibility guidance row. Dictionary, Settings, diagnostics, and Metrics dashboard rows expose presenter-backed accessible names for screen readers.
Primary and secondary recording shortcuts each support key-based or modifier-only Toggle, Push to Talk, and Hybrid modes. Windows uses a low-level key-up hook only for recording shortcuts and keeps utility shortcuts on RegisterHotKey.
While the floating recorder is active, Ctrl+1 through Ctrl+0 selects prompt slots and Alt+1 through Alt+0 selects enabled Power Mode rules.
Power Mode rules can also send a post-insertion key after successful dictation insertion: Return, Shift+Return, or Windows-adapted Ctrl+Return for macOS commandEnter compatibility. Enabled rules can define direct global shortcuts that select that rule immediately, using the same recorder capture format as the Settings shortcut fields. The Power Mode editor validates rule behavior before saving: enabled non-default rules need at least one process, title, or browser URL match, multiple enabled defaults are rejected, and duplicate or ignored match fields are shown as inline warnings.
About / Open Source includes local diagnostics actions for opening the app data folder, copying a safe diagnostics summary, and exporting a .log diagnostic report with system/app context, known file metadata, and recent generic in-app status labels. Diagnostic exports do not include API keys, Credential Manager values, environment variables, clipboard contents, transcript/history text, rendered AI prompt contents, user-authored dictionary/prompt text, or settings file contents.
Settings Recording Feedback controls cover local start/stop sound feedback, custom start/stop sound import, muting the default Windows output while recording, opt-in media pause/resume while recording, and the shared resume delay. Audio/media restore runs after capture stops so long transcription work does not keep output muted. Stop sound plays after successful text insertion. By default, sound feedback uses Windows system sounds; use Choose to import a local .wav, .mp3, .aiff, or .aif sound up to 3 seconds for either Start Sound or Stop Sound. Imported sounds are copied under %LocalAppData%\VoiceInk.Windows\Sounds, can be tested from Settings, can be reset back to System Default, and fall back to Windows system sounds if the custom file cannot be played. Media pause/resume is off by default and uses Windows Global System Media Transport Controls, pausing only when Windows reports a current playing session and resuming only that same session when supported. During long transcription, enhancement, or insertion work after capture stops, Cancel Recording cancels the active processing pipeline cooperatively, keeps VoiceInk responsive, and preserves the captured audio as a canceled History row when possible.
For cloud transcription, choose OpenAI-compatible in AI Models, then pick a provider preset. The preset area shows macOS-style provider metadata for language coverage, speed, accuracy, description, and whether Windows currently supports batch transcription or realtime recorder preview for that provider. The custom preset lets you enter a full request endpoint such as https://api.example.com/v1/audio/transcriptions and model ID. Groq, Deepgram, AssemblyAI, Mistral, ElevenLabs, Soniox, Speechmatics, Gemini, xAI, and Cartesia fill provider-specific endpoints and model choices. Deepgram, AssemblyAI, Soniox, Speechmatics, and Cartesia can stream interim recorder preview text when Show Live Transcript Preview is enabled. API keys are stored per preset in Windows Credential Manager. Test Provider sends a metadata-only key probe for Custom OpenAI-compatible, Groq, Mistral, xAI, Deepgram, AssemblyAI, ElevenLabs, Soniox, Gemini, Speechmatics, and Cartesia; it does not upload audio and does not include response bodies or API keys in status text. Gemini uses inline audio for short recordings and automatically switches to the Gemini Files API for larger WAV recordings before the final generateContent transcription request. Remote endpoints must use HTTPS; plain HTTP is accepted only for localhost development endpoints. VoiceInk rejects endpoint URLs that contain embedded credentials or common key/token query parameters so API keys stay out of JSON settings. Dictation, Transcribe Audio, and Retry Last then use that provider instead of local Whisper, sending provider-appropriate audio transcription requests with vocabulary prompt context when supported. For credential-gated manual smoke testing, use docs/superpowers/windows-cloud-provider-smoke-runbook.md.
- WinUI 3 shell with sidebar navigation for Dashboard, Transcribe Audio, History, Metrics, AI Models, Enhancement, Power Mode, Permissions, Audio Input, Dictionary, Settings, and About / Open Source
- Windows tray icon with show/hide, recording toggle, quick-setting submenus, Quick Add, History, and Quit commands
- Always-on-top floating recorder with Mini and Notch styles, state text, elapsed timer, live microphone level bars, non-activating Stop/Cancel controls, hover-dismissable no-activate Prompt/Power chooser panels, a gated live transcript preview panel for real partial sources, and pulse animation
- About / Open Source replacement for commercial VoiceInk Pro surfaces, with local-only diagnostics folder, summary, and diagnostic log export actions
- First-run setup dialog for local model path, microphone settings/input, primary shortcut, basic usage, and explicit UI Automation names for programmatic setup controls
- Permissions section with shortcut, microphone, text insertion, and screen context readiness cards plus Windows microphone privacy routing
- Local Whisper catalog cards with direct GGML
.bindownloads, app-local model storage, imported.binreferences, selected model path health checks, stale imported model cleanup, model-aware transcription language selection, nonblocking warmup/preload, and default model selection from the shell - Local whisper.cpp transcription through Whisper.net
- Default-off OpenAI-compatible cloud transcription provider with custom, Groq, and Deepgram presets, endpoint/model settings, endpoint-query
response_formatsupport for OpenAI-compatible requests, provider-specific Windows Credential Manager API key storage, multipart OpenAI-compatible requests, direct Deepgram batch requests with advanced endpoint query options, Deepgram live recorder preview streaming, sanitized HTTP errors, and routing across Dictation, Transcribe Audio, and History Retry - Transcribe Audio section with multi-file picker and drag-and-drop import, restored pending/failed queue snapshots, pending/processing/completed/failed queue states, cancel/clear/remove/retry/copy/save/enhance controls, Media Foundation import to app-owned WAV recordings, selected-provider transcription, cleanup, and History save
- Default-off AI Enhancement section with Custom/Cerebras/Groq/Gemini/OpenAI/OpenRouter/Mistral/Ollama provider presets, OpenAI-compatible and native Ollama endpoint/model settings, provider-specific Windows Credential Manager API key storage, automatic read-only selected text context, optional read-only Clipboard Context, macOS-aligned Default/Assistant/Chat/Email/Rewrite prompts, Assistant context wrapping, prompt selection and local custom prompt editing, trigger-word persistence, timeout/short-phrase/retry settings, output filtering, endpoint safety validation, and original-text fallback
- Dictation, Transcribe Audio, and history retry can run enhancement after local transcription cleanup and before insertion when enhancement is enabled and configured
- Power Mode section with ordered enabled/default rules, active-window quick fill from Win32 foreground-window process/title detection, direct per-rule shortcuts, session-only overrides for model/language/enhancement/prompt/cleanup settings, post-insertion auto-send keys, and history name/emoji metadata
- Microphone capture
- Clipboard-based text insertion into the active app
- JSON settings and SQLite transcription history
- Configurable global key+modifier or modifier-only shortcuts for primary and secondary recording with Toggle, Push to Talk, and Hybrid modes, plus key-based paste last, paste last enhanced, retry last transcription, cancel recording, open history, and quick add to dictionary
- Core dictionary models for vocabulary words and word replacements
- Persistent JSON-backed Dictionary storage for vocabulary words and word replacements
- macOS-style transcription cleanup for hallucination markers, filler words, punctuation cleanup, lowercase output, trailing spaces, and word replacements
- Vocabulary prompt biasing for local Whisper transcription
- Expanded history metadata for original text, final text, enhanced text, status, language, model path, prompt name, Power Mode, enhancement provider/model, enhancement timing, AI request messages, and audio file path
- Shell controls for filler-word removal, punctuation cleanup, lowercase output, and trailing spaces
- Settings Clipboard controls for restore delay and default/direct-text paste methods
- Settings General
Launch at Logincontrol using the current user's Windows startup registry key - Settings Privacy controls for macOS-style transcript cleanup, audio-only cleanup, and confirmed onboarding reset
- Shell controls for adding/removing/sorting vocabulary words and adding/editing/enabling/disabling/removing/sorting word replacements
- Quick-add Dictionary dialog for vocabulary words and word replacements
- Dictionary JSON import/export for vocabulary words and word replacements
- Shell audio input refresh plus endpoint-ID backed System Default, Custom Device, and Prioritized microphone selection
- Shell recent-history list/detail view for original, final, enhanced, status, timing, model, prompt, and error metadata
- Picker-based CSV history export
- Paste-last final and enhanced-preferred history actions in the shell
- History search, accessible History list row names, selected-row audio playback/open, selected-row retry, retry-last-to-clipboard, active-recording cancel history, and confirmed single-item delete
- Settings JSON backup export/import with category selection for General Settings, Custom Prompts, Power Mode, Dictionary, and Custom Model Definitions; Credential Manager API keys are intentionally excluded
The current shell does not yet expose richer visual parity for Settings/Metrics/Power Mode/Dictionary. The first-run setup dialog, onboarding reset, sidebar navigation, Permissions readiness page, About / Open Source diagnostics actions, local Whisper catalog downloads/imported model selector/language selector/warmup, broad OpenAI-compatible and provider-specific cloud transcription path, Deepgram/AssemblyAI/Soniox/Speechmatics/Cartesia live recorder preview, tray command surface with quick-setting submenus for model/provider/enhancement/language/audio/context/Power Mode settings, floating recorder Mini/Notch styles with live microphone level bars, non-activating Stop/Cancel controls, hover-dismissable Prompt/Power chooser panels, Ctrl/Alt digit mini-recorder shortcuts, and gated live transcript preview rendering for real partial sources, Enhancement settings/default prompt/provider/custom-prompt/selected-text/clipboard/OCR/browser-context/native-Ollama path, Power Mode process/title/browser URL/default/recorder-selection/shortcut/auto-send rule path, Transcribe Audio picker/drag-drop restored queue/copy-save/per-file-enhance/provider-history path, Dictionary add/edit/sort/import/export/quick-add path, Settings backup import/export, Settings privacy cleanup controls, recent-history metadata view, paste-last actions, retry-last action, active-recording cancel action, open-history focus action, selected-row history retry/playback/open/re-enhance, history search/delete/export, batch history actions, configurable global key+modifier shortcuts with recording key-up modes including modifier-only primary/secondary recording shortcuts, and endpoint-ID backed System Default/Custom/Prioritized microphone selection are source-runnable and wired into the dictation pipeline. Installer packaging is part of the Windows fork scope and is planned as a follow-on Windows subsystem after this source-built MVP.
This project is not accepting pull requests at this time. You're welcome to fork and modify VoiceInk for your own use.
You can still contribute by:
- Reporting bugs via issues
- Suggesting features or enhancements
- Improving documentation via issues
For more details, see our Contributing Guidelines. For build instructions, see our Building Guide.
This project is licensed under the GNU General Public License v3.0 - see the LICENSE file for details.
If you encounter any issues or have questions, please:
- Check the existing issues in the GitHub repository
- Create a new issue if your problem isn't already reported
- Provide as much detail as possible about your environment and the problem
- whisper.cpp - High-performance inference of OpenAI's Whisper model
- FluidAudio - Used for Parakeet model implementation
- Sparkle - Keeping VoiceInk up to date
- KeyboardShortcuts - User-customizable keyboard shortcuts
- LaunchAtLogin - Launch at login functionality
- MediaRemoteAdapter - Media playback control during recording
- Zip - File compression and decompression utilities
- SelectedTextKit - A modern macOS library for getting selected text
- Swift Atomics - Low-level atomic operations for thread-safe concurrent programming
Made with ❤️ by Pax