Speak where you usually type.

The complete app is free on GitHub under the MIT License. The Microsoft Store and Mac App Store versions are paid; each Store shows its current price. They offer the same core features and updates through the Store. The Mac App Store version omits the AppleScript paste fallback and the permission repair tool. The app requires neither a subscription nor a WhisPaste account. Optional cloud providers may charge separately for usage.

For Windows, Apple Silicon Macs and Linux. Intel Macs are not supported. Linux packages are available as AppImage, .deb and tar.gz for x64. The download page links to the installers and AppImage; all packages are also available in GitHub Releases. WhisPaste is a desktop app, with no iOS or Android version.

At least 8 GB of RAM and a 64-bit processor: Windows 10 or later, macOS 12 Monterey or later on Apple Silicon, or Linux. You also need disk space for the app and models. A GPU is optional. For Whisper, start with Balanced from about 2 GB of graphics memory on NVIDIA or Apple Silicon, or about 4 GB on AMD or Intel; below that, or without a GPU, choose Quick & Compact. Apple Silicon uses shared memory. The optional local Smart Mode model needs about 3 GB of extra disk space and around 4–5 GB of RAM while running. Auto-Paste must be enabled: macOS requires Accessibility permission, and Linux needs write access to /dev/uinput. On Windows, an app running with elevated permissions can prevent pasting. Protected or unsupported fields can block Auto-Paste.

For downloads outside the Microsoft Store, SmartScreen may warn when a file has little download reputation. Check that your download came from the official WhisPaste repository. For this type of SmartScreen notice, you can continue through “More info” → “Run anyway”. The paid Microsoft Store version is another option.

The current GitHub DMG is signed and notarized by Apple. The paid Mac App Store version is distributed through Apple. Both are for Apple Silicon only. If macOS blocks an older or self-built copy, System Settings → Privacy & Security may offer "Open Anyway".

Local speech recognition with Whisper or Parakeet works offline after a one-time model download. Local Smart Mode needs an additional model download, and the optional CUDA pack needs additional files. Cloud speech recognition and Smart Mode through OpenAI require an internet connection. Model downloads, update checks and enabled statistics or crash reports can use network connections.

With local speech recognition, audio stays on your computer and is discarded after transcription unless you enable audio retention. Regular transcripts are kept in local history; quick notes are stored separately in Notes. Cloud speech recognition sends audio directly to OpenAI or Deepgram, while Smart Mode through OpenAI sends text. Usage statistics and crash reports are on by default and can be turned off separately in Settings → Privacy. Statistics contain no audio, transcripts or history. File paths in crash reports are not currently redacted.

Choose Whisper for its 99 languages or choice of model sizes. Parakeet covers about 25 European languages and offers fast transcription even without a GPU. It was faster than the compact Whisper model in our CPU benchmark; results depend on the hardware and test conditions. Both engines support GPU acceleration and work offline after the model download. Switch in Settings → Speech Recognition.

Whisper supports 99 languages; Parakeet supports about 25 European languages. Select the language of your recording in Settings or use automatic detection. This is separate from the app’s interface language. Translation into seven target languages is a separate, optional Smart Mode feature.

Recognition depends on the language, recording and model. For Whisper, Balanced is a good starting point from about 2 GB of graphics memory on NVIDIA or Apple Silicon, or about 4 GB on AMD or Intel. Below that, or without a GPU, choose Quick & Compact. Larger models can help with difficult audio but need more memory and processing time. Review the result before sending; you can change models later.

Smart Mode is off by default. Cleanup removes filler words and adjusts punctuation, Concise shortens the text, and Translate converts it into one of seven target languages: German, English, Spanish, French, Portuguese, Mandarin or Russian. Apply a preset during recordings or later in your history. Local processing works on macOS and Windows and is experimental on Linux. Alternatively, use your own API key to send the text directly to OpenAI. Review the wording and meaning of revised text.

Save the text as a snippet. Enable the snippet hotkey or configure a trigger word for the picker. Speak that word on its own in a recording; the picker opens after transcription. The hotkey opens it without recording. Text is inserted only when you select it. Interactive snippets guide you through a template’s fields and then insert the completed text.