Free open-source community edition

Talk to your computer.
Keep the words local.

Whisper Dictate Community Edition is a local push-to-talk dictation app for Windows and macOS. After the selected speech model is acquired, transcription runs on your machine and does not upload your audio.

View community downloads Read the source

No account, payment, or analytics is required to use this static site. Model weights are not included in the app; their publisher licenses apply.

Local by designAudio stays on your computer
Model-awareWeights download on first use
Open developmentInspect and build from source
Ctrl+Shift+SpaceEnglish dictation Ctrl+Alt+SpaceGerman dictation

What “community edition” means

This page is for the free, open-source community build. It does not sell access, promise a subscription, or claim that every future artifact is available yet. The repository’s current license and the separate model licenses are the authoritative terms; read them before redistributing software or weights.

A small, inspectable dictation toolkit

The community source documents the workflow below. Exact behavior and availability can change with the source revision and your configuration.

Push to talk

Hold a configurable hotkey, speak, release, and insert the transcription at the focused text field.

Local post-processing

Deterministic rules can remove filler words, apply spoken punctuation, and format simple structure.

English and German

The documented profiles use separate NVIDIA Nemotron streaming models and language-specific configuration.

Shortcuts and commands

Text snippets and a small command mode are available in the source for users who want repeatable phrases and edits.

Your own hardware

The published Windows preview ships a CPU-only PyTorch build, so transcription runs on the processor. The macOS app uses Apple GPU acceleration where the hardware supports it.

Auditable source

Read the implementation, configuration, and setup instructions directly in the source repository.

Release manifest

Community downloads

View releases.json

Community preview release 0.1.0 is published through the canonical GitHub release. The Windows portable ZIP is explicitly unsigned; the macOS Apple Silicon ZIP is Developer ID signed and notarized. Installer buttons remain disabled until verified installer packages are available. There is no guessed or placeholder download endpoint.

Loading the release manifest…

Portable ZIP or installer

Windows x64

Windows 10/11 · 64-bit · CPU-only build; this preview does not bundle a CUDA runtime, so an NVIDIA GPU is not used.

Portable ZIP

Extract and run from a writable folder. No system-wide install is implied by this option.

Installer

A guided installer will appear here only after a verified community installer is published.

The current Windows portable ZIP is an unsigned preview; its SHA-256 is published with the release manifest.

Portable ZIP or installer

macOS Apple Silicon

macOS 13+ · Apple Silicon (arm64) · Apple GPU acceleration where supported; CPU fallback is documented.

Portable ZIP

Unzip the app bundle and launch it from your Applications folder or another writable location.

Installer

An Apple Silicon installer is not advertised until its package and checksum are verified.

The current macOS portable ZIP is Developer ID signed and notarized. Intel macOS remains outside this release track.

Prefer to build from source?

Clone the source repository, follow its setup instructions, and use the documented Python environment. A source build is not the same thing as a verified packaged artifact.

Open source setup

First run

Acquire once, then work offline

“Offline” applies after model acquisition. The community build does not bundle the model weights, so the first use of each language needs a network connection.

  1. 1

    Choose a verified artifact

    Use a link from the release manifest, then compare its SHA-256 checksum with the manifest before launching it.

  2. 2

    Start the app once online

    At startup the app fetches the documented NVIDIA model for every configured language. It tries the configured mirror chain first and falls back to Hugging Face, so a first run still works when huggingface.co is unreachable. Each checkpoint is currently described as approximately 2.4 GB.

  3. 3

    Review the model terms

    The model publisher’s license applies to the weights. They are downloaded directly by the app and are not redistributed with this site or the app archive.

  4. 4

    Use local transcription

    Once the model is cached, recording and transcription stay local. Later dictation does not need to send audio to a speech service.

Separate model licenses

Speech models are not part of the community license

Read license notices

The model publisher supplies the weights and sets the terms. This site links to the model cards and licenses; it does not mirror or relicense the weights.

Model download privacy: the first model request contacts the configured mirror host or Hugging Face and may reveal your IP address and request metadata to that service. The app does not upload recorded audio. After a model is cached, the local transcription path does not require that download connection.

System requirements

Requirements can differ by source revision and artifact. Treat the manifest and release notes as the source of truth for a published build.

Windows x64

  • Windows 10 or 11, 64-bit
  • Microphone and keyboard access
  • CPU-only build in this preview (no bundled CUDA runtime)
  • A CPU-only transcription is slower than a GPU build

macOS Apple Silicon

  • macOS 13 or newer, arm64
  • Microphone permission for the launcher
  • Accessibility / Input Monitoring for dictation and paste
  • Apple GPU acceleration where supported; CPU fallback is documented

Storage and source builds

  • Allow roughly 2.4 GB per downloaded language model
  • Leave additional space for the app, cache, and logs
  • Packaged artifacts do not require Python
  • Source setup documents Python 3.12+

Verify before you run

Checksums confirm that a downloaded file matches the manifest. They do not replace reviewing the release notes or source.

SHA-256 checksums will appear here when a canonical release is published.

PowerShell

Get-FileHash .\WhisperDictate-….zip -Algorithm SHA256

macOS / Linux

shasum -a 256 WhisperDictate-….zip

If a checksum differs, do not run the file. Re-download it and compare the exact filename and hash from releases.json.

Troubleshooting

Start with the release manifest and the source README. The answers below describe the documented community workflow, not a support SLA.

Why are the download buttons disabled?

No canonical community release has been entered into releases.json yet. The site intentionally refuses to invent a download URL. Watch the source repository’s release page and update the manifest only after artifacts and checksums are verified.

Why does the first run need internet?

Speech model weights are not bundled. The first use of each language downloads the selected model from Hugging Face. After the model is cached, the documented transcription path runs locally and can work offline.

What permissions does macOS need?

Grant Microphone access to the app or launcher. Dictation and text insertion also use Accessibility / Input Monitoring, and the documented launcher may request Automation access for System Events. Denied permissions can allow recording without allowing text to be inserted.

What if transcription is slow or the GPU is not used?

Check the release/source configuration and the app log. The published Windows preview is a CPU-only build, so an NVIDIA GPU is not used there; the macOS app uses Apple GPU acceleration where the hardware supports it. A CPU-only run is generally slower than an accelerated one.

What should I do if a checksum does not match?

Do not run the artifact. Confirm the exact filename, download it again from the canonical URL in the manifest, and compare the full SHA-256 digest. A mismatch is a reason to stop and investigate.

Can I redistribute a ZIP or the models?

Do not assume that the model or software terms permit redistribution. Read the repository software license, the linked model licenses, and the repository’s third-party notices first.

Build it yourself

The source repository is the reference

Read the code, setup instructions, configuration, and license before using or redistributing a build. This site is a separate community landing page and does not change the repository’s license or model terms.

Open GitHub source Read software license Third-party notices