Voicetypr and offline dictation that inserts at your cursor

The whole loop: shortcut, record, transcribe, insert

Voicetypr is an open-source desktop dictation app for macOS and Windows, described as offline-first voice-to-text dictation. The pitch in one line is that you speak in any app and Voicetypr transcribes your voice and places the result at your cursor.

The quick start in the documentation is five steps, and they are worth reading because they describe the insertion model rather than a separate editor. Open the app and choose a local or cloud transcription model. Set the primary recording shortcut in General settings. Place the cursor in any text field. Press the shortcut, speak, and stop recording. Voicetypr inserts the transcript at the cursor and stores it in local history.

The feature list spells out the input options: system-wide shortcuts, push-to-talk, toggle recording, and automatic insertion at the active cursor. That is a different design from a dictation app that opens its own window and makes you copy the result out. Cursor insertion is also the part with the sharpest platform requirements, which is why Accessibility permission on macOS is called out in the installation requirements.

The history is a real feature rather than a log. You can search it, filter it, inspect metadata, compare original and formatted text, copy it, save it, or re-transcribe it.

Whisper on both platforms, Parakeet on Apple Silicon

Local transcription is the default, and the engines are named. Whisper runs on macOS and Windows, and there are Apple Silicon-optimized Parakeet models on macOS, delivered as a sidecar process.

The distinction matters for a Mac user. Apple Silicon is the primary macOS target and supports both Whisper and Parakeet, while Intel packages remain available as a legacy Whisper-only download while the project evaluates usage before a separately announced end-of-life. That phrase about a separately announced end-of-life is a clear signal that Intel Mac support is on the way out, and anyone on an Intel machine should factor that into the decision.

Windows is treated differently. It can use the bundled CPU transcription path on every supported machine, and optional Vulkan acceleration runs in an isolated sidecar process and falls back to CPU if the GPU path is unavailable. The isolation is the point: the architecture table in the documentation says the main executable remains CPU-safe, which means a GPU driver problem cannot take the application down with it.

The model files are downloaded once and stored locally. That is a one-time cost in bandwidth and disk, and it is also what makes the local mode genuinely offline rather than a thin client in disguise.

Four modes, and a table that tells you exactly what leaves the machine

Voicetypr is offline-first, but the documentation is careful that the selected mode determines what leaves your computer, and it publishes a data flow table rather than a slogan.

In local transcription, recorded audio and the transcription stay on the device, with model files downloaded once and stored locally. In cloud transcription, the recorded audio is sent to the cloud speech-to-text provider you selected, and the optional providers listed are Soniox, OpenAI, Groq, Deepgram and Cohere. In AI formatting, the transcript is sent to the AI provider you configured for rewriting, with OpenAI, Anthropic, Gemini or a custom OpenAI-compatible endpoint. In Network Sharing, audio is sent to the Voicetypr server you explicitly configured.

That last mode is unusual. You can use another Voicetypr installation as a private transcription server on your LAN or configured network, with the architecture describing it as an authenticated Voicetypr server and client pair. It is a way to give a machine without a capable GPU access to one that has a local model, without routing audio through a third party.

Diagnostics and product-analytics controls are available in Settings, and the documentation points at a privacy policy for the current collection and retention details. File transcription is also supported, with supported cloud providers offering speaker diarization.

The CLI that makes the app scriptable

The most underrated feature is the agent-ready CLI, described as giving scripts and local AI agents audio-to-text and microphone capture with plain-text or structured JSON output.

You install the voicetypr command from Settings, then Advanced, and then use it from a terminal or an agent:

voicetypr --help
voicetypr status --json
voicetypr models --json
voicetypr transcribe --file note.wav --json
voicetypr record --until-silence --json

Human-readable output is the default, and --json switches it to structured output for automation. The privacy property holds here too: audio stays on the machine unless the command explicitly selects a remote Voicetypr server.

This changes what the application is. A dictation app is a personal tool, but a transcription CLI is a component you can wire into a script, a note-taking pipeline, or an agent that needs to capture a thought and write it down. The commands cover both directions of the problem: transcribe a file you already have, and record until silence from the microphone.

The commercial framing follows the same logic. With a lifetime license and a local model, the documentation states there are no per-minute API fees or cloud usage quotas, while optional cloud modes still use the selected provider's billing and limits.

Building from source: Tauri, Rust, React and two test suites

The build is a Tauri v2 project, so the prerequisites are Node.js and pnpm, a Rust stable toolchain, the Tauri v2 platform prerequisites for your operating system, and Xcode command-line tools on macOS or Visual Studio Build Tools on Windows.

The clone and dev sequence is:

git clone https://github.com/ideaplexa/voicetypr.git
cd voicetypr
pnpm install
pnpm tauri:dev

The verification commands are the usual pnpm names, and the split is informative:

pnpm lint
pnpm typecheck
pnpm test
pnpm test:backend
pnpm check

test runs vitest on the frontend and test:backend runs cargo test inside src-tauri, so the Rust core and the React surface are tested separately. There is a test:all script that runs both, and a benchmark script backed by a perf-harness.mjs file, with a perf-corpus directory at the repository root for the input files.

The layer breakdown in the documentation explains the shape of the code. Tauri v2 handles windowing, menus, updater integration, permissions and packaging. The frontend is React 19 with TypeScript, Tailwind CSS, shadcn/ui and Zustand. The backend is Rust handling recording, resampling, transcription orchestration, hotkeys, history and cursor insertion. The local engines are Whisper on macOS and Windows, with the Parakeet sidecar on Apple Silicon.

The repository also carries AGENTS.md and CLAUDE.md with repository conventions and architecture notes, plus a sidecar directory at the root matching the Parakeet and Vulkan sidecars described in the architecture.

AGPL-3.0, two distribution channels, and an Intel deadline

The licence is AGPL-3.0, and that choice has a consequence for the Network Sharing mode that is worth thinking about before you use it. The network clause of the AGPL covers software offered to users over a network, which is exactly what a Voicetypr server acting as a transcription backend for other machines is. If you run one of these servers for other people, you are in the part of the licence that is stricter than MIT or Apache-2.0. This is a reading of the licence, not legal advice, and the LICENSE file is the text that governs.

Distribution comes in two shapes on Windows. There is a direct installer updated through Voicetypr's Stable or Beta channel, and a Microsoft Store listing which is updated through the Store, with the documentation noting that Microsoft Store installations remain Store-managed. The release channels feature lets you choose Stable or Beta updates, and the repository contains MICROSOFT_STORE_LAUNCH.md alongside a build-msix-store.ps1 script for the Store packaging path.

The release history shows both a stable line and a parallel beta line: v2.0.9 and v2.0.8 on 2026-10-09, and v2.1.0-beta.4 on 2026-10-03, with a changelog published at voicetypr.com/changelog. The repository is not archived and the last push was on 2026-10-16.

Two support statements are worth holding onto. macOS requires macOS 14 or later, with Apple Silicon fully supported and Intel described as legacy and best-effort, and release builds are signed and notarized by Apple. Windows requires 64-bit Windows 10 build 19041 or later, or Windows 11. The Intel path is the one to watch.

Editorial conclusion

Voicetypr is a good fit for a developer or writer who dictates sensitive content, works on a laptop without reliable connectivity, or wants an audio-to-text command for local agents, because the local path keeps audio on the machine and the CLI exposes the same pipeline to scripts. It is a poor fit for a Linux user, for a team that needs shared transcripts with speaker attribution at scale, or for anyone unwilling to install a model and accept a trial before the lifetime license. Before adopting, confirm your machine clears the stated floors, macOS 14 with Apple Silicon preferred or 64-bit Windows 10 build 19041, and read the licensing terms carefully, since AGPL-3.0 covers a network-facing server in a way most desktop licences do not.

Frequently asked questions

Official sources

1. ideaplexa/voicetypr on GitHub
2. License: AGPL-3.0
3. Project website
4. README
5. Releases