Choose your platform, finish setup, and try an in-app dictation with harmless words before using Presspeech in another app. Speech is transcribed on-device; model downloads use the network. The in-app scratchpad is not isolated from the system clipboard.
Downloading is not launching: macOS 0.3.8 and Windows 0.1.12 start a missing-model request when Presspeech opens, not when you download the ZIP or installer. Neither macOS 0.3.8 nor Windows 0.1.12 lets Setup defer a missing-model request after launch; macOS 0.3.9 and later ask before a clean install’s first download. Unsure about a token, telemetry, or proxy? Keep it closed.
Safe stopping point: You can download a build and leave it unopened; installing without opening it also does not start the request. If you install Windows but wait, clear the installer’s final Launch Presspeech option. It starts checked by default; confirm it is clear before choosing Finish. Before choosing to open it, read the macOS decision or Windows decision.
Waiting with an existing install? Quit any running copy. Without opening Presspeech again, turn off its entry, if enabled, in macOS System Settings → General → Login Items & Extensions or Windows Settings → Apps → Startup. On Mac, also deselect Reopen windows when logging back in at sign-out or restart. Check GitHub Releases in a browser rather than opening the old app to look for updates; follow the new build's instructions before launching it. These steps cannot undo a model request already started.
Choose macOS 0.3.9 after reviewing its launch decision if a Hugging Face token may be inherited by Presspeech, or you are unsure. Do not launch 0.3.8 while using a TLS-inspecting proxy whose trust is unclear.
Download on launch: a missing model is about 500–600 MB. A model request can include the inherited token; the bundled client honors lowercase https_proxy, and a TLS-inspecting proxy trusted by macOS can read that token.
Proxy URL risk: 0.3.8 can log a malformed inherited https_proxy or http_proxy URL verbatim, including credentials in that URL, then ignore the proxy. If you depend on one and are unsure its URL is valid, leave 0.3.8 unopened. From 0.3.9, Presspeech refuses malformed proxy settings before model loading; it still honors valid proxies.
Read the full macOS warning before deciding to launch.
Wait for published 0.1.13 if you want to avoid possible Hugging Face usage telemetry, an available token or custom download route may be configured, or you are unsure. If you install but wait, leave Launch Presspeech unchecked.
Download size and language: on a fresh PC with usable NVIDIA CUDA, the missing default model is about 2.5 GB and multilingual; without usable CUDA, the CPU default is about 141 MiB and English-only. Check language support before downloading.
Routing risk: A custom route can change where the request and token go; a TLS-inspecting proxy trusted by the client can read the token. If such a proxy's trust is unclear, do not launch while it is in use.
Published 0.1.12 also briefly opens the selected microphone for an automatic local readiness check in first-run Setup; samples are discarded, not saved or sent. If you do not want that check, leave 0.1.12 unopened. Upcoming 0.1.13 waits for an explicit Check Microphone choice.
Read the full Windows warning before deciding to launch. The upcoming build disables bundled-library telemetry and account-token authentication but still honors proxy and CA settings.
The public models need no account token, and model requests do not include dictation audio or transcripts. Do not inspect or share token values. Windows 0.1.13 is not published yet; check GitHub Releases before following a newer guide. Source can be ahead of downloads. If you choose to launch now, read your platform's full warning and the version-specific network inventory.
These steps describe the macOS 0.3.8 and 0.3.9 releases and Windows 0.1.12 prerelease, not features planned for later downloads.
If a step fails, stop there. Use its guidance below or match the visible status to the recovery guide.
Requires Apple Silicon and macOS 14 or later. The release is notarised. On an existing installation, a missing 500–600 MB local model starts downloading on launch; a clean install asks first, and a valid cached model loads without another download. Speech recognition runs on your Mac.
Already used macOS 0.3.8? Read the follow-up guidance before another model download.
Requires x64; Windows 11 is recommended. Use Windows 10 only while your specific edition remains supported or is covered by Extended Security Updates; general support ended on 14 October 2025.
Allow about 4.4 GB for the app plus a first model download: usable NVIDIA CUDA selects multilingual Parakeet (~2.5 GB); without usable CUDA, the CPU default is English-only Whisper base.en (~141 MiB). A missing selected model starts downloading on first launch. See the full language and model guide.
The installer is unsigned. Follow the SHA-256 verification steps. After checking the source and hash, you may choose More info → Run anyway for a SmartScreen warning. If Smart App Control or managed policy blocks it, stop—do not try to circumvent the block.
Already used Windows 0.1.12? Read the token-safety guidance for a past model download. Do not include token values in logs or support requests.
Already installed? After checking the first-launch decision, continue to finish setup.
Not supported: Intel Macs and Windows on Arm. For Windows 10, follow the support qualification above; see the Windows install guide before downloading.
A missing model begins downloading when you launch 0.3.8; in 0.3.9 and later, a clean install waits until you choose Download Model. Keep Presspeech open while it prepares, and reopen Setup Checklist from the menu-bar item if needed. Continue when the model reports Ready.
Use its buttons for Microphone, Accessibility (called Device Control and Data Access on macOS 27 and later), and Input Monitoring. Accessibility is a broad system-control grant: Apple advises granting it only to apps you trust. Presspeech uses it to check the focused destination and send the paste shortcut; without it, the normal verified paste path is unavailable.
Before granting Input Monitoring: macOS's grant can expose typed keys across apps. Presspeech uses keyboard events to detect the configured hotkey and Escape to cancel recording; it passes other keys through without saving, logging, or sending their values. Read Apple's Input Monitoring guidance and the full permission scope before deciding.
Hold and release the displayed key. The default is Right Option. If it does not respond or controls another feature, choose a different key in Settings. You can open Try Dictation once Presspeech is Ready, before this test succeeds, and use the menu's Start Dictation and Stop and Transcribe controls. The checklist says Done only after a real hotkey event.
Wait until local model preparation finishes and Try Dictation becomes available. In 0.1.12, Set Up Later does not stop a missing-model download that began on launch.
0.1.12 checks the selected microphone automatically at startup and when its selection changes; choose Check Again to repeat it. The check briefly opens the selected input to measure its level; samples stay in memory and are discarded, not saved, sent, or transcribed. Windows may show its microphone-use indicator while the input is open. Speak during the check; if no level appears, select or unmute the correct input and repeat Check Again.
The default is Right Alt. If it acts as AltGr, choose F8 or another available key. On a new profile, 0.1.12 selects Start Presspeech with Windows by default; turn it off before choosing Finish Setup, Set Up Later, or closing Setup if you do not want the app at sign-in. Its Setup has no dictation-style selector: choose Press to toggle in Settings. Review these choices, then finish setup. See the version-specific first-launch steps.
Use a harmless practice sentence in the built-in scratchpad to check the complete flow without sending text into another app.
Before you practice: the scratchpad is an in-app destination, not a guarantee that the system clipboard is untouched. macOS 0.3.8 can put dictated text on the clipboard during delivery; Windows 0.1.12 can copy it for recovery if focus changes. Manually copying scratchpad text also uses the clipboard. With Handoff enabled, macOS Universal Clipboard may make copied text available on nearby Apple devices; Windows Clipboard History or sync and third-party clipboard managers may retain it. Use only non-sensitive test words, then review the clipboard privacy boundary before dictating anything sensitive.
On macOS, choose Support → Try Dictation…. On Windows, choose Try Dictation… from setup or the notification-area menu.
Leave that field focused so Presspeech can verify the delivery path.
Hold the displayed key, say a short sentence, then release. Or choose Press to toggle—under Settings → Dictation → Trigger on macOS, or in Windows Settings—to press once to start and again to stop. On macOS you can also choose Start Dictation and Stop and Transcribe from the menu. Continue when your sentence appears. Audio is discarded after local processing; the scratchpad is not a saved history.
No recording cue? Recheck the hotkey and microphone in Setup Checklist on macOS or Settings on Windows. Do not retry while the model is still preparing.
Open a blank TextEdit document on macOS or Notepad on Windows, click in it, and repeat hold–speak–release without changing windows. Keep that field selected until dictation finishes; a window-level check does not reliably establish the identity of a field or browser tab. Press Escape to cancel an active recording.
In every new target app, start with a blank, disposable field that cannot submit or execute text. Dictate harmless words once. Count automatic paste as working only if the complete text appears exactly once in the intended field without a recovery notice. A successful scratchpad or TextEdit/Notepad test does not establish automatic-paste support in another app. Use the result guide below before pasting or recording again.
If the complete sentence appears in the intended field, this is an automatic-paste smoke pass for this field only. Test another app separately before relying on it.
This is manual recovery, not automatic paste. Check the field first. Before pasting, verify the current clipboard only if safe; a later copy may have replaced the transcript.
Stop and inspect the disposable field before trying again. A delivery error does not prove the field is empty or that the clipboard holds the transcript. Follow the macOS or Windows recovery steps.
Keep the original field and browser tab selected until Presspeech finishes. A switch within the same window may paste into the newly focused place without a copied/manual-paste notice. If you changed focus, inspect the destination before retrying; use only disposable fields for a target-app test.
If Presspeech did not paste: check the intended field first; text may already be there. A macOS 0.3.8 copied notice or Windows 0.1.12's Transcript copied, not pasted means the finished transcript was copied at completion, not that the clipboard still holds it for manual paste. A later copy can replace it without changing the notice. Other delivery-error notices do not guarantee that the clipboard holds the transcript or that the field is empty; do not paste or dictate again blindly. Match the notice to the macOS or Windows recovery steps.
If unsure and it is safe to inspect the current clipboard item, paste into a blank, disposable field in a local, non-syncing, non-executing text editor you trust—not a web form, chat, or command shell. Paste into your intended field only if it still holds the complete transcript; do not dictate the same text again before checking. If the clipboard may hold sensitive unrelated content, do not paste it just to check. On macOS, if the clipboard changed, use Copy Last Transcript when Recent Transcripts is enabled. Otherwise, do not paste the newer clipboard item as your dictation.
This is recovery, not a successful automatic paste. It can happen in some Electron/Chromium apps on macOS 0.3.8 even when the window appears unchanged. On Windows, reopen an administrator-level target normally rather than elevating Presspeech.
If macOS 0.3.8 reports “Couldn't paste”: inspect the field before retrying. An input error does not prove nothing arrived; text may be partial or duplicated. If text is missing or incomplete, remove any partial text before using Copy Last Transcript when offered to paste the complete transcript deliberately. If no recovery copy is available, correct or remove partial text before another recording. Do not use that target for important text until you have checked repeatable delivery with harmless phrases. See the paste recovery guide.
Command shells can run pasted text. Append newline, Windows newline, or spoken formatting can submit text. Dictate commands into Try Dictation or a text editor, review the exact text, then paste and run deliberately. See Apple and Windows terminal guidance.
One disposable-field try is a smoke check, not proof of repeatable delivery. Before relying on a browser editor, Electron app, terminal, or remote session, use the eight-attempt privacy-safe target-app check. It separates automatic paste from manual recovery without publishing your words.
If you hid the item in Settings → Behavior → Show Presspeech in Menu Bar, right-click the Dock icon and choose Show Presspeech in Menu Bar. Hiding the item enables Dock access if needed. If a crowded or notched menu bar temporarily omits it, open Presspeech.app again from Applications, Finder, or Spotlight; the running app opens Setup Checklist instead of starting another copy. Enable Show in Dock there if you need a persistent alternative.
Check the notification-area overflow first. You can also launch Presspeech again from the Start Menu to restore its window or controls without starting a second dictation process.
For model, permission, audio, or hotkey recovery, use the troubleshooting guide. If the problem remains, Copy Diagnostics copies a privacy-safe summary but does not send it; review it and keep it private until there is a usable report route. Never post dictated text, audio, references, or dictionary contents to a public issue.
Public issue creation is restricted (checked 23 September 2026). The in-app Report a Problem and Suggest an Improvement actions open the support guide; they do not submit anything. Copy Diagnostics copies a summary locally and does not send it. GitHub currently blocks creating new issues, so do not expect a new report form to accept a submission. Check existing issues for a relevant thread and whether comments are available. If there is no usable route, keep your report draft private and retry later; do not post private data elsewhere as a workaround. For a security concern, follow the private reporting instructions.
Keep only details useful for reproduction; this draft is not sent or monitored. Record the Presspeech and operating-system versions, affected stage, expected and observed result, smallest safe reproduction steps or frequency, generic target-app context if relevant, recovery tried and whether it worked, and attempt/failure counts. For text delivery, use harmless test text and record only whether complete text was recovered—never save or share the phrase itself. Keep diagnostics private until reviewed; do not include audio, transcripts, raw logs, credentials, private paths, or identifying names.