Voice to text app for Windows, inspired by VoiceInk for macOS. Transcribe what you say, almost instantly, locally or through the cloud.
Note: Parla is a Windows port / re-implementation inspired by VoiceInk by Pax. All Swift / macOS-specific code is re-written from scratch in Rust + TypeScript. VoiceInk remains the behavioral reference - every feature decision in Parla is cross-checked with its VoiceInk counterpart.
Parla is a native Windows app that transcribes your voice to text and pastes it at your cursor, with a full pipeline that goes far beyond raw transcription:
- Accurate transcription via local Whisper (whisper.cpp), NVIDIA Parakeet (ONNX Runtime), or cloud providers.
- Privacy-first by default - you can run the whole pipeline (transcription + LLM enhancement) 100% offline.
- Power Mode - detects the foreground window and browser URL, then applies a pre-configured profile (which model, which LLM prompt, which dictionary) automatically.
- Context aware - the foreground window is captured + OCRed, and the extracted text is injected into the LLM enhancement prompt so the output matches what you were working on.
- Global shortcuts with configurable toggle, push-to-talk or hybrid mode, plus a double-Escape to cancel a recording.
- Personal dictionary of word replacements applied on every transcription (technical terms, product names, custom acronyms).
- Prompt detection via trigger words in your transcript - say "mail ..." and the Email prompt kicks in automatically.
- Multilingual UI - French, English, Spanish.
- NVIDIA CUDA acceleration for Whisper, Parakeet and local llama.cpp, with transparent CPU fallback.
- Auto-updater via GitHub Releases, so the app keeps itself in sync with new releases.
Compared to existing Windows dictation tools, Parla targets feature parity with VoiceInk on macOS:
- Accurate transcription with local (Whisper / Parakeet) or cloud (Groq, Deepgram, ElevenLabs, Mistral, Soniox, Speechmatics, Gemini, Custom OpenAI-compat) models.
- Streaming transcription for providers that support it (ElevenLabs Scribe v2, Deepgram nova-3, Mistral voxtral RT, Soniox stt-rt-v4, Speechmatics).
- LLM enhancement with an embedded llama.cpp (GGUF models with CUDA), Ollama local, plus all major cloud providers (Anthropic, OpenAI, Gemini, Mistral, Groq, Cerebras, OpenRouter) and local CLI templates (pi / claude / codex) or a custom PowerShell command.
- Smart modes (prompts) configurable per-profile: Default cleanup, Assistant, Email, Chat, Rewrite, or any custom prompt you define.
- Multi-format clipboard backup - if you had an image, a file, or rich text in your clipboard, it is restored byte-for-byte after paste.
- Non-focus-stealing overlay recorder pill that floats above all windows without capturing focus (Win32
WS_EX_NOACTIVATE+WS_EX_TOOLWINDOW). - History stored in SQLite with CSV export and configurable retention.
- Windows Credential Manager for every API key - nothing is ever written to disk in plain text.
| Source | Models | GPU acceleration |
|---|---|---|
| whisper.cpp (local) | tiny, base, small, medium, large-v2, large-v3, large-v3-turbo, large-v3-turbo-q5_0, custom .bin import |
CUDA / cuBLAS |
| Parakeet (local) | parakeet-tdt-0.6b-v2 (EN), parakeet-tdt-0.6b-v3 (25+ langues) | CUDA EP, DirectML EP, CPU |
| VAD Silero | v5.1.2 (via whisper-vad) |
CPU |
| Groq | whisper-large-v3-turbo | - |
| ElevenLabs | scribe_v1, scribe_v2 realtime | - |
| Deepgram | nova-3, nova-3-medical | - |
| Mistral | voxtral-mini-latest, voxtral-mini-transcribe-realtime-2602 | - |
| Soniox | stt-async-v4, stt-rt-v4 | - |
| Speechmatics | speechmatics-enhanced | - |
| Google Gemini | 2.5 Pro, 2.5 Flash, 3 Flash, 3.1 Pro | - |
| Custom | Any OpenAI-compatible endpoint | - |
| Provider | API key | Endpoint |
|---|---|---|
| llama.cpp embedded | No | In-process (GGUF models, CUDA optional) |
| Ollama local | No | http://localhost:11434 (configurable) |
| Anthropic | Yes | https://api.anthropic.com/v1/messages |
| OpenAI | Yes | https://api.openai.com/v1/chat/completions |
| Google Gemini | Yes | https://generativelanguage.googleapis.com/v1beta |
| Mistral | Yes | https://api.mistral.ai/v1/chat/completions |
| Groq | Yes | https://api.groq.com/openai/v1 |
| Cerebras | Yes | https://api.cerebras.ai/v1 |
| OpenRouter | Yes | https://openrouter.ai/api/v1 |
| Local CLI (PowerShell) | No | Local process (pi / claude / codex templates or custom) |
| Custom OpenAI-compat | Optional | User-configured URL (HTTPS enforced) |
- Windows 10 (22H2) or later
- A microphone
- (Optional) NVIDIA GPU with CUDA 12.x drivers for GPU-accelerated local transcription and LLM enhancement
Head to the releases page and grab Parla_x.y.z_x64-setup.exe. Works on every Windows 10 / 11 machine. Runs Whisper, Parakeet and llama.cpp on your CPU out of the box.
Auto-updater is built in: once installed, Parla checks latest.json on every start and offers to install new releases in place.
The pre-built installer ships CPU-only because compiling whisper.cpp + llama.cpp + ggml-cuda does not fit in our public CI runners' budget. Users with an NVIDIA GPU who want CUDA acceleration (typically 3 to 10 times faster on large models) build Parla themselves from source - see BUILDING.md for the recipe. You will need the CUDA Toolkit 12.x installed plus Visual Studio 2022 with the C++ workload. The build takes 15 to 30 minutes on a reasonably recent desktop.
Once installed, Parla runs in the system tray. The default global shortcut for recording is configurable from the Recorder panel in the app.
See BUILDING.md for the detailed instructions. Quick version:
git clone https://github.com/LitteRabbit-37/Parla.git
cd Parla
npm install
npm run tauri devFor a release build:
npm run tauri build- Tauri v2 desktop framework (Rust backend + WebView2 frontend)
- React 19 + TypeScript + Vite + Tailwind v4 + shadcn/ui for the UI
- whisper.cpp via
whisper-rsfor local Whisper transcription - parakeet-rs (ONNX Runtime + TDT decoder) for local Parakeet transcription
- llama.cpp via
llama-cpp-2for local LLM enhancement - cpal for WASAPI audio capture
- xcap for screen / window capture + Windows
Media.Ocrfor OCR - uiautomation for Chromium / Firefox / Edge URL extraction
- rusqlite (bundled) for the history database
- keyring with Windows Credential Manager backend for API keys
- react-i18next for the UI translations (French, English, Spanish)
src-tauri/src/
audio/ - WASAPI recorder + VAD Silero
transcription/ - engine (VoiceInk parity), pipeline, Whisper, Parakeet
cloud/ - batch + streaming providers
enhancement/ - LLM orchestrator + providers + prompt detection
power_mode/ - foreground window detection + URL extraction + profile matcher
screen_context/ - window capture + OCR
history/ - SQLite + retention + CSV export
paste/ - Ctrl+V via SendInput + multi-format clipboard backup
hotkeys/ - low-level keyboard hook + state machine
src/
components/ - UI panels + mini recorder overlay
lib/tauri.ts - IPC bindings
i18n/ - locale files (fr, en, es)
Parla uses a global low-level keyboard hook to detect its recording shortcut (Right Alt by default). Windows UIPI (User Interface Privilege Isolation) prevents a non-elevated process from receiving keyboard events destined for a window of higher integrity level. In practice, this means that when a terminal, an editor, or any other application has been started with Run as administrator, the Parla hotkey is silently swallowed while that elevated window has keyboard focus, and no recording starts.
The proper fix is to ship Parla with a manifest flag uiAccess=true, which Windows honors on code-signed binaries installed under C:\Program Files. We do not currently have an Authenticode code-signing certificate, so this is not available yet.
Workarounds until we sign the releases:
- Use a non-elevated terminal whenever possible. Windows Terminal, PowerShell and Command Prompt only need administrator rights for a very small set of commands.
- If you genuinely need an elevated terminal, run Parla as administrator as well (right-click
parla.exe> Run as administrator). Both processes then share the same integrity level and the hotkey is delivered. Note that an administrator Parla cannot paste into non-elevated applications - UIPI enforces the rule symmetrically. - Give focus to a non-elevated window before triggering the hotkey, then switch back.
This limitation is a property of Windows, not a bug in Parla. It will be removed in a future release when signed installers are available.
- BUILDING.md - Detailed build instructions
- CONTRIBUTING.md - How to contribute
- CHANGELOG.md - Release notes
- Windows 10 22H2 or later (Windows 11 recommended for best
Media.Ocrlanguage coverage) - At least 4 GB of RAM for the base Whisper model, more for the larger ones
- (Optional) NVIDIA GPU with CUDA 12.x drivers for GPU acceleration
Bug reports, feature requests, and pull requests are all welcome. See CONTRIBUTING.md for details.
You can also help by:
- Reporting bugs or regressions via issues
- Suggesting features or enhancements
- Improving the documentation or translations (FR / EN / ES)
This project is licensed under the GNU General Public License v3.0. See the LICENSE file for details.
Parla is an independent project inspired by VoiceInk (GPL-3.0) but does not share any Swift / macOS source code. The Windows implementation is a full re-write in Rust + TypeScript.
- whisper.cpp - High-performance inference of OpenAI's Whisper model
- parakeet-rs - Rust wrapper over ONNX Runtime for NVIDIA Parakeet TDT
- llama.cpp - Inference engine for the embedded LLM
- Tauri - The desktop framework that ties everything together
- VoiceInk by Pax - the macOS reference this project aims to match feature-for-feature
- FluidAudio - the macOS Parakeet implementation VoiceInk relies on
- whisper-rs - Rust bindings over whisper.cpp
- llama-cpp-rs - Rust bindings over llama.cpp
- cpal - Cross-platform audio I/O
- keyring-rs - OS-level credential storage
- rusqlite - SQLite bindings for Rust
- tokio-tungstenite - async WebSocket for streaming providers
If you find Parla useful and want to support the development, you can buy me a coffee:
Many thanks to Pax for the original VoiceInk and to every contributor of the upstream projects (whisper.cpp, llama.cpp, parakeet-rs, Tauri, shadcn/ui).
Feel free to submit issues, pull requests, or feedback.