Private push-to-talk dictation for Mac and Windows.

Hold a key, speak, and release—or choose Press to toggle. Presspeech transcribes locally, with no account or cloud transcription. Before opening a current build, review the model-download decision below.

macOS 0.3.10: released, signed, and notarised for Apple Silicon on macOS 14+. Windows 0.1.12: an unsigned prerelease for x64 PCs; Windows 11 is recommended.

Before opening a current build

Unsure? Leave it unopened. Downloading the ZIP or installer does not start a speech-model request; installing without opening Presspeech does not either; first launch with a missing model does, and neither macOS 0.3.8 nor Windows 0.1.12 lets Setup defer that request (macOS 0.3.9 and later ask first on a clean install). On Windows, clear the installer’s final Launch Presspeech option if you choose to wait. It starts checked by default; confirm it is clear before choosing Finish.

  • macOS 0.3.8 — inherited token or proxy URL: A missing-model request may include a Hugging Face token. If one may be inherited, or you are unsure, choose macOS 0.3.9 after reviewing its launch decision. Do not launch through a TLS-inspecting proxy whose trust is unclear. A malformed inherited proxy URL may also be logged verbatim, including any embedded credentials, and ignored; if you depend on a proxy and are unsure its URL is valid, leave 0.3.8 unopened. Read the full macOS launch decision.
  • Windows 0.1.12 — telemetry or token: A missing-model request may send Hugging Face usage telemetry or an available token; custom routing can change its destination. First-run Setup also briefly opens the selected microphone for an automatic local readiness check. If you want to avoid possible telemetry, a token or custom route may be configured, or you are unsure, wait until Windows 0.1.13 is published. Do not launch through a TLS-inspecting proxy whose trust is unclear. Read the full Windows launch decision.

Review the version-specific network-call inventory before choosing to launch.

Check the macOS release Check the Windows prerelease

Already installed? Follow the four-checkpoint first-dictation guide.

Before dictating into another app: Keep the starting field and tab selected until delivery finishes. The published builds verify the original window, not a field or browser tab in the same window. If Presspeech cannot safely verify the destination, it avoids automatic paste. Check before recovering: On macOS 0.3.8 and published Windows 0.1.12, a copied notice means the transcript was copied at completion, but a later copy can replace it. Other delivery-error notices do not guarantee that the clipboard holds the transcript or that the field received nothing. Check the intended field first; paste manually only if the clipboard still holds the complete transcript. See the first-dictation guide for recovery steps.

~100 msmodel transcription on benchmark clips
9.2 MBsigned release zip
~80 MBRAM while idle
0% CPUbetween dictations
Choose your platform

Choose the build that matches your computer

macOS — released

The signed and notarised Swift menu-bar app runs multilingual Parakeet TDT v3 through CoreML on Apple Silicon. The 0.3.8 release starts a missing 500–600 MB local model download on launch, while 0.3.9 and later ask first on a clean install; a valid cached model loads without another download.

Before opening macOS 0.3.8: a missing-model request may include a Hugging Face token inherited by Presspeech; an inherited lowercase https_proxy can route it through a TLS-inspecting proxy trusted by macOS. If a token may be present—or you are unsure—choose macOS 0.3.9 after reviewing its launch decision. If the proxy's trust is unclear, do not launch 0.3.8 while it is in use. Downloading the ZIP does not start the request, so you can leave the app unopened. The public model needs no token; dictation audio and transcripts are not sent. Read the full macOS warning before deciding to launch.

Already used macOS 0.3.8? See what to do before another model download.

Review macOS requirements, privacy, and install steps

Windows — prerelease

The currently unsigned x64 installer is for evaluation on Windows 11 (recommended) or Windows 10 only while your specific edition remains supported or is enrolled in applicable Extended Security Updates. Microsoft ended general support for Windows 10 on 14 October 2025. Allow about 4.4 GB for the app, plus a first model download of about 2.5 GB with usable NVIDIA CUDA (multilingual Parakeet) or 141 MiB on CPU (English-only Whisper base.en). Published 0.1.12 starts a missing selected model download on launch. Other local models remain selectable.

Before opening Windows 0.1.12: model downloads may send Hugging Face usage telemetry and an already-configured or locally saved token. Custom routing can change the request destination; a TLS-inspecting HTTPS proxy trusted by the client can read a token it receives. If you want to avoid possible telemetry, a token or custom route may be configured, or you are unsure, wait until 0.1.13 is published. If a TLS-inspecting proxy's trust is unclear, do not launch while it is in use. Downloading the installer and checksum does not start a model request; if you install but wait, leave Launch Presspeech unchecked. The public models need no token; dictation audio and transcripts are not sent. Read the full Windows warning before deciding to launch.

Already used Windows 0.1.12? If a model download ran with a token available and an inherited HF_ENDPOINT or staging setting may have sent it to a destination you do not trust—or an untrusted TLS-inspecting HTTPS proxy could read the request—treat the token as disclosed to that destination or proxy. A proxy that only tunnels HTTPS cannot read the token. Revoke the token and create a replacement at Hugging Face Access Tokens; do not include token values in logs or support requests.

Windows may block an unsigned installer, especially under Smart App Control or managed policy. Review the verification and security guidance before downloading; do not bypass a policy block.

Review Windows requirements, privacy, and verification steps

Current published downloads: macOS 0.3.10 and Windows 0.1.12 prerelease. Install metadata updated .

The whole workflow

Hold. Speak. Release.

The video below shows the macOS build with its default Right Option hotkey. On Windows, use the same hold–speak–release gesture with Right Alt by default; if that key acts as AltGr for your keyboard layout, Setup lets you choose F8 or another available key. Press to toggle is available in macOS Settings → Dictation → Trigger and in Windows Settings.

Silent visual walkthrough: Right Option is held while Presspeech shows that it is listening, then released. After local transcription, the complete sentence appears in the same TextEdit document. If Presspeech cannot verify the original window, the published build copies the transcript at completion for manual recovery; paste only if the clipboard still holds it. A field or browser tab switch inside that window may go undetected, so keep the starting field and tab selected.

Check an app before relying on automatic paste. Target apps, including browser editors, Electron/Chromium apps, terminal-based editors, and remote sessions, expose focus and paste behavior differently. On macOS 0.3.8, some Electron/Chromium apps can show the copied/manual-paste notice even when focus looks unchanged; that is recovery, not an automatic-paste pass. Run the privacy-safe target-app check to separate automatic paste from clipboard recovery; it records outcomes, never your words.

New in macOS 0.3

Say exactly what should happen

Voice shortcuts

Map a phrase you say to exact reusable text. No rewriting model, no cloud call, and no ambiguity about the result.

Spoken formatting

Opt in to deterministic English commands including “new paragraph”, “bullet point”, “comma”, and “open quote”; selecting the French Language Hint switches formatting to canonical French phrases.

Try Dictation

Both apps include an in-app scratchpad for a first test. It may still use the system clipboard, so practice with harmless words. Use the first-dictation guide to decide whether to launch the current build before testing the hotkey.

macOS implementation

Built for one thing, not ten

Press and speak

Presspeech is a menu-bar utility, not a document manager. The default trigger is Right Option, with other hotkeys available from the menu.

Local speech recognition

Audio is captured in memory and transcribed locally. Presspeech checks the original window before automatic paste. A copied notice means the published build copied the transcript at completion for manual recovery; check that the clipboard still holds it before pasting. Other delivery errors do not guarantee a copy or an empty field. Audio is then discarded and transcripts are not written to disk.

Native macOS stack

Swift, AVFoundation, CoreGraphics, AppKit, CoreML, and FluidAudio. No embedded browser runtime, interpreter, account, subscription, or cloud transcription endpoint.

Good fit

When Presspeech is the right pick

Use Presspeech when you want

Fast phrase-length dictation in everyday text fields, with a visible clipboard fallback when automatic delivery is unavailable; a local-only transcription model; and a workflow that does not keep a transcript library. The macOS app is released and notarised; the Windows app is available for people comfortable evaluating an unsigned prerelease.

Use something else when you need

Intel Mac or Windows on Arm support, macOS before 14 Sonoma, an installer accepted by managed Windows policies, batch file transcription, speaker labels, long-form transcript editing, cloud models, team features, or integrations with notes databases.

Why trust it

Evidence and install paths

BenchmarkedFluidAudio on the Apple Neural Engine is measured against the prior GPU path on the same WAV clips.
AuditableThe macOS app is a single Swift source file; the separate Windows implementation is readable Python. Release scripts and packaging metadata are public.
Private by designThe version-specific network inventory covers model downloads and bundled-library behavior, optional update checks, and user-triggered support and install links.
Delivery is testableThe eight-check target-app protocol distinguishes automatic paste, safe clipboard recovery, and incorrect results without publishing anyone's words.
Agent-readableInstall instructions, benchmark results, privacy facts, and comparison pages are available as plain HTML, Markdown, JSON, and text.

Feedback availability checked 24 September 2026: GitHub currently restricts new issues. The target-app worksheet can save aggregate results as a local, unmonitored draft; downloading it does not submit a report. Test for your own decision, then check the current support route before sharing counts. If no comparable open thread accepts comments, keep the draft private until a suitable route reopens.

Test one target app Browse compatibility reports