Test Presspeech with one app

Separate insertion from recognition

Presspeech binds each recording to the window where it began. It should paste only when it can still verify that destination. A focus change should avoid insertion and leave a route to deliberate manual recovery, but the clipboard state is version-specific. macOS and published Windows 0.1.12 normally leave a copy on the clipboard; Windows 0.1.13 and later can keep it only in memory until Delivery Recovery → Copy for Manual Paste is chosen. On macOS, if Presspeech detects another copy replacing the clipboard during delivery, it stops and preserves that newer copy; 0.3.8 shows Couldn't paste, while builds with revised wording say Delivery uncertain. A failed input event does not prove that no text reached the destination, so inspect the target before retrying or using recovery. After a copied notice, a later copy can replace the transcript without updating that notice; use Command-V only if the clipboard still holds the dictation, or use Copy Last Transcript when available. Revised macOS copy notices make this condition explicit.

macOS 0.3.8 checks the original window, not the field or browser tab within it. The 0.3.9 implementation also compares the focused Accessibility control when an app exposes one at recording start. If no control is exposed then, or two fields reuse one identity, a same-window move can still send the transcript to the new focus. Keep the original field and tab focused until insertion or a recovery notice appears. The separate-window checks below do not establish safety for same-window moves.

This is a delivery check, not a transcription-quality test. The recognized wording can differ from what was spoken and still pass: the question is whether the same finished transcript reaches the intended field once, or is recovered safely without reaching another field.

Choose a useful target

If you already depend on one app, test that app. Otherwise, choose a safe disposable field in a target class that does not yet have a comparable report. The live searches below make missing coverage visible without collecting anything on this site.

Before testing

1. Finish Presspeech setup

   Read the published-build privacy check above before launching or downloading a model. Use the current official build. On macOS, this baseline requires 0.3.8 or later; update before testing if the About window shows 0.3.7. Confirm the model is ready, the microphone and hotkey checks pass, and Try Dictation works first.

2. Open two disposable fields

   Use one blank field in the target app and one blank field in a local scratch app. Never use a production message, document, account, command shell, or field where pasted text can submit, send, or execute. Test a remote/virtual environment or terminal-based editor only when it is isolated, disposable, and non-executing; otherwise skip that target.

3. Use harmless, distinct phrases

   Create a different phrase for every attempt so stale or duplicate delivery is visible. Do not use names, credentials, commands, customer data, or text from real work.

   Optional eight-phrase prompt

   If English suits the selected language hint, use these in order: amber rabbit one; blue otter two; copper robin three; green badger four; ivory falcon five; purple fox six; silver heron seven; yellow turtle eight. Otherwise use the same colour-animal-number pattern in the test language. Recognition accuracy is not scored; compare the finished target text with the scratch paste. Never add either version to the public report.

4. Make clipboard behavior comparable

   On macOS 0.3.8 and later, leave Keep Previous Clipboard for Manual Restore off for this baseline. Published Windows 0.1.12 leaves the finished transcript on the current clipboard. Windows 0.1.13 and later leaves the previous item unchanged for an already-known invalid target; do not assume a recovery notice means the transcript is on the clipboard. Its actual dictation clipboard writes are marked for exclusion from Windows Clipboard History and Cloud Clipboard; published 0.1.12 does not mark them. macOS 0.3.8 can expose transcript entries to macOS Universal Clipboard; disable Handoff if the harmless test text must stay on the Mac. Builds containing local-only transcript clipboard writes prevent Universal Clipboard transfer while preserving local Command-V. Apple's Screen Sharing has a separate Edit → Use Shared Clipboard setting; those transcript writes are not a tested guarantee against that remote transfer. Turn shared clipboard off if you do not intend to transfer test text to the remote Mac. If macOS Clipboard History in Spotlight is enabled on macOS 26 or later, clear it after the check if you do not want the text retained there. Also disable third-party clipboard managers if you do not want the test text retained elsewhere.

5. Check the target's own Paste shortcut

   Copy a harmless marker by ordinary means, then use physical Command-V on macOS or Control-V on Windows once in that same blank target field. Confirm it appears exactly once and clear the field before dictating. This is a target-app shortcut baseline, not a Presspeech result or one of the eight worksheet attempts. If it fails, do not score that target as a Presspeech paste pass or guess that Presspeech caused the failure; keep a local note of the target version and try another safe target. A working context-menu Paste does not establish that the keyboard shortcut works.

6. Record versions, not private context

   Note the Presspeech, operating-system, and target-app versions. Classify the target as native, browser/web editor, Electron/Chromium, terminal, remote desktop/virtual machine, or elevated Windows app, and record the generic field type (for example, plain text, rich text, or browser content editor). Note the active keyboard layout/input source only if it differs from your usual layout or helps explain an outcome; record its public name, never typed content. Omit account, server, document, tab, form, and window names.

Keep the target focused

Plan five attempts with a different harmless phrase each time. Stop early if an attempt is incorrect or unsafe; do not repeat it just to fill the worksheet.

1. Start in the blank target field

   Keep its window focused from the start of recording until Presspeech finishes. Do not copy anything or allow a clipboard tool to replace its contents during delivery.

2. Dictate once, then inspect the target

   Note whether one copy appeared and whether Presspeech showed a version-appropriate recovery notice.

3. Compare the finished text using the right recovery route

   First inspect the target field. After an automatic paste with no recovery notice, paste the current clipboard into the separate local scratch field before copying anything else and compare the complete text, including its suffix. After a macOS copied notice or published Windows 0.1.12's Transcript copied, not pasted, check the clipboard in the scratch field before manual paste. The notice reflects a copy at completion, not a guarantee that a later clipboard change left the transcript current. After Windows 0.1.13 and later's Delivery Recovery notice, do not treat the current clipboard as the transcript: choose Copy for Manual Paste explicitly, then paste that copy into the scratch field. The prior clipboard item may still be current. On macOS, a Couldn't paste or Delivery uncertain warning also does not prove the clipboard contains the transcript or that the target received nothing; inspect the target, then use Copy Last Transcript if offered. Remove any partial target text before pasting a full recovery copy. If no recovery copy is available, record a failed delivery; deliberate recovery never turns an attempt into an automatic-paste pass.

4. Classify one outcome

   Use the definitions below, then clear both disposable fields before the next attempt.

Not completed means an attempt was not run. If you cannot continue for another reason, classify each unrun slot that way; do not call a completed failure not completed. An early stop is not a passing five-attempt baseline.

If you know another copy changed the clipboard during delivery, and it prevented classification without an incorrect or unsafe result, record it as an interruption under Relevant conditions and repeat with a fresh harmless phrase if it remains safe to continue. Exclude only those safe, known, unclassifiable interruptions from the five scheduled check slots; count completed pasted-once and recovered-safely outcomes. A stale, partial, duplicate, misdirected, or unrecoverable result is Incorrect or unsafe even when the clipboard change explains it: count that completed attempt and stop. Never exclude an unexplained failure.

Move focus before dictation finishes

Use blank disposable fields in two separate windows. Start dictation in the first window, move focus to the second, then finish. In hold mode, move before releasing the hotkey; in toggle mode, move before the stop press.

Classify each slot as copied for manual paste without inserting anywhere, inserted into a field, no insertion, but recovery failed, or not completed. The first category includes an explicit, verified Copy from Windows 0.1.13 Delivery Recovery; it does not imply the app copied automatically. Recovery failed means a completed attempt inserted nowhere but the notice or complete transcript was unavailable; it is an incorrect result, not an unrun check. Do not exclude insertion or failed recovery as an interruption merely because another copy changed the clipboard. The four counts should total three.

An older seven-count draft with five classified steady-focus attempts can add Steady-focus not completed: 0. An older six-count draft may combine failed recovery and unrun focus-change slots as Other or not completed. If that count is zero, both new focus-change counts are zero. If it is nonzero, do not guess or present the draft as a comparable eight-count baseline. Keep it private and, if safe, rerun the protocol using the current categories.

Check a tab switch separately

The eight checks use separate windows; they cannot establish safety when a browser switches tabs inside one window. The macOS 0.3.8 implementation checks only the window, so this is already a known limitation of that build. Do not repeat it to claim a new 0.3.8 safety result. If you have a macOS 0.3.9 or later build with focused-control checks, this optional observation can help identify a browser/version that still needs native qualification.

Run it only after an eight-check baseline with no unsafe result. Use two blank, non-submitting fields in two tabs of the same browser window, and harmless disposable text. First verify ordinary physical Command-V in both fields and clear them; skip this observation if either shortcut fails. In toggle mode, start one dictation in the first tab, switch to the second before stopping, and keep it focused until Presspeech finishes. If switching safely while recording is not possible, skip the observation. Inspect both fields before copying, pasting, or retrying. Do not switch back before delivery or test in a real form, chat, or command window.

Record categories, not words

Use this optional worksheet as you complete the checks above. It records only the outcome category selected for each check slot, including unrun slots, calculates the two sets of aggregate counts, and prepares the exact count and overall-result block used by the compatibility report. There is nowhere to enter a phrase, transcript, or report context.

Keep platform boundaries separate

Share counts, not content

1. Find the same platform, app, version, and field type

   Browse existing compatibility reports. When the platform, app version, and generic field type match, add a comment there instead of opening a duplicate—even if your outcome differs. Paste the worksheet block, then add the Presspeech and operating-system versions, generic hardware if useful, and relevant conditions, because variation under comparable conditions is important evidence. Open a separate report for a different platform or field type.

   Comment shape for another observation

   Presspeech version: [x.y.z]
   Operating-system version: [version]
   Hardware (optional, no serial or device names): [generic model/chip]
   
   [paste the worksheet's eight counts and Overall result]
   
   Relevant conditions: [trigger mode, keyboard layout/input source when
   relevant, suffix, clipboard manager/history, assistive technology, or a minimal
   reproduction; omit private context]
   Optional remote clipboard handoff observation (if safely run, separate from
   eight counts): [client/version, local and guest OS, sharing state, attempts,
   anomalies or recovery notices, final scratch comparison; no text or hostnames]

2. If GitHub accepts new reports, submit one target-app report

   If no matching report exists and GitHub currently accepts submissions, open the compatibility report form. Paste the worksheet's report-ready block into Eight-check outcome counts and select the same overall result in the form, or enter the same values manually. For a new report, put the platform, public app name/version, and generic field type after the fixed [Compatibility]: title prefix. Passing results document successful delivery or safe recovery for that exact tested configuration; self-selected reports cannot establish a population success rate. If issue creation is restricted, keep the draft locally and retry later rather than repeatedly opening the blocked form.

   If issue creation is still restricted and your macOS Electron result directly addresses issue #33, you may instead comment there if GitHub permits it. Include the eight counts plus Presspeech, operating-system, and target-app versions and the generic field type. Do not post unrelated results in that thread; retain those drafts privately when no suitable route accepts them.

   For Overall result, a complete automatic-paste pass means all five steady-focus attempts pasted once and all three focus-change attempts recovered safely. Focus-change recovery is expected; choose the manual-recovery option only when it occurred while the original target stayed focused. Any incorrect or unsafe result takes precedence.

3. Review diagnostics before pasting

   Copy Diagnostics produces a privacy-safe summary if you choose to include it; exact microphone names, raw error details, and raw local log lines are not copied into that report. Never add dictated or recognized text, clipboard contents, audio, dictionary or shortcut contents, screenshots with user data, credentials, titles, names, or private paths.

Community reports are exploratory evidence for the exact versions and conditions recorded. They do not replace native release qualification or promise universal support.