The Secretary

The Secretary turns a spoken thought into text where you are working. Press your hotkey, say what you mean, and carry on. Start with a free local engine; try another when the job calls for it.

Think of it as a small writing desk, not another writing project. The six speech engines share one microphone, vocabulary list, history, and hotkey. Their strengths differ; your first successful sentence should not require learning all of them.

The app is now called The Secretary. Older betas used Transcriber or Concierge. Existing saved data, export folders, and some macOS permission entries may retain an older name. Use the app's version in Preferences → About to identify the copy you are running; do not delete old settings simply to tidy the name.

This guide covers the Mac 1.0 beta (build 4). The iPhone and iPad chapter describes the companion development version, which does not yet match every Mac feature. Experimental features may change before release.

First time here? Follow Quick start on Mac. If text appears twice, go straight to Duplicate text. If words are missing, keep the recording and see Recording length.

Contents#

1. Overview
2. Requirements
3. Quick start on Mac
4. Choosing a transcription engine
5. Hotkeys
6. The Secretary window and menu
7. Preferences
8. Translation, enrichment, and automation
9. Purchase, licensing, and provider charges
10. iPhone and iPad
11. Privacy and stored information
12. Troubleshooting
13. Reference links
14. Glossary
15. Revision history

Overview#

The normal Mac workflow is deliberately simple:

1. Place the insertion point in any editable field in any app.
2. Start recording with the configured speech-recognition hotkey.
3. Speak. Secretary's live preview is hidden by default so you can concentrate on the thought.
4. Stop recording. Secretary processes the recording locally or with the selected cloud service. GPT Live streams audio while you speak instead of waiting for the end.
5. With the default settings, the completed text is pasted once, kept on the clipboard, and saved in Secretary's history. Test Window, Watch Folder, cancelled or failed recordings, and sessions with automatic paste turned off do not paste into the active app.

Choose an engine in Preferences → Providers before the next recording. One global hotkey controls the engine named in the green LIVE badge. The badge means selected engine, not microphone currently listening. The recording status tells you when capture is active. There is no Done button for ordinary preference changes.

Secretary is not the same thing as Apple's system Dictation or Voice Control. They can coexist, but each has its own shortcut, vocabulary, settings, and recognition behavior.

Requirements#

Mac#

• macOS 13 Ventura or later.
• A working built-in, USB, or Bluetooth microphone.
• Microphone permission for every engine.
• Speech Recognition permission when using Apple Speech.
• Accessibility permission for modifier-only hotkeys such as Hold Fn, Hold Option, Hold Control, or Double-Tap Option. Some custom Fn combinations may also require it.
• Internet access for Gemini, Meta Muse, GPT Transcribe, GPT Live Transcribe, target-language translation, AI enrichment, and the initial Local Whisper model download.
• A provider API key for Gemini, Meta, or OpenAI. Apple Speech and Local Whisper do not need API keys.
• About 145 MB for the Fast Local Whisper model or about 483 MB for the More Accurate model, plus additional space macOS may use while preparing an optimized Core ML copy.
• Enough free space for Local Whisper models and any recoverable audio. Optional plaintext export uses ~/Documents/Transcriber/Transcriptions/ only when you explicitly enable it.

Apple Speech can work offline only when the selected locale supports on-device recognition and the necessary speech resources are installed. Secretary checks this before it starts an Apple recording.

iPhone and iPad#

• iOS or iPadOS 16 or later.
• Microphone permission, plus Speech Recognition permission for Apple Speech.
• Internet and provider credentials for cloud engines.
• Allow Full Access for the optional Secretary keyboard extension when it needs shared settings and provider access.

Recording length#

There is no built-in “two sentences” rule for Whisper. Nor is there one dependable maximum that covers every Mac, model, language, microphone, and cloud account. Missing a paragraph is a problem to investigate, not normal behavior to accept.

Cost & History → Dictation Length Science shows measured completed recordings, longest completed duration, processing time, and warnings. For Local Whisper, diagnostics also describe how much audio was analyzed and where recognized text begins and ends. A completed duration is evidence of that attempt, not a promised safe limit. Failed attempts are not a complete part of the aggregate study yet.

For an important long passage, first run a controlled test in Test Window:

1. Say “Blue volcano begins this recording.”
2. Read an ordinary paragraph without exaggerated syllables. Say “Silver pineapple marks the middle,” then read another paragraph.
3. End with “Golden compass ends this recording.” Stop and check that all three markers and both paragraphs are present.
4. Compare the captured duration and any warnings in Log. If text is missing, keep the recoverable WAV and listen to it before changing settings. That separates missing audio from recognition or delivery problems.

These markers are example test phrases, not special commands. Build up the length gradually on your own setup. For a meeting or irreplaceable material, keep an independent source recording and transcribe a copy through Watch Folder. Successful dictation audio is normally removed; History is text, not an audio backup.

Installing and updating the Mac app#

Download the free Mac beta, version 1.0 (build 4). This is a Universal app for Apple silicon and Intel Macs, signed by Plum Amazing Software LLC and notarized by Apple. Optional cloud-provider charges are separate; Apple Speech and Local Whisper do not require a paid API account.

Open the ZIP, quit other running Secretary/Transcriber betas, move The Secretary.app to Applications, and launch that copy. Check Preferences → About for version 1.0 (4). Grant permissions only when macOS asks for them. Do not run multiple copies to compare engines; use Test Window instead.

If macOS blocks the application, verify that it came from Plum Amazing and contact support rather than disabling system security. Existing preferences normally survive an app replacement. Keep your recovered audio and exported text when upgrading; neither is a substitute for a normal Mac backup.

Quick start on Mac#

1. Launch Secretary and grant the requested Microphone permission.
2. Open Preferences → Test Window.
3. Choose an engine. Apple Speech is the quickest no-account test; Local Whisper is private after its one-time download; cloud engines require their API keys.
4. Click the Test Window microphone, speak, and click again. The result stays in Test Window and is not pasted into another app.
5. Choose a hotkey and read the complete arrow-separated instruction beneath it.
6. Click Arm Hotkey Test, perform the gesture, speak, and stop using the same gesture. Confirm that Test Window reports both the received hotkey and a transcription result.
7. Click End Hotkey Test.
8. Place the insertion point in another app and use the same hotkey. The completed text should paste at the cursor.

New Mac installations select Apple Speech, Control + Option + Space — Press to Start/Stop, Smart formatting where supported, Paste final, automatic paste, clipboard copy, and sounds. Translation, automatic enrichment, plaintext exports, and Watch Folder trashing are off. A saved preference from an earlier beta can differ.

Your first useful sentence#

With Apple Speech selected, leave optional translation and enrichment off. In Test Window, say: “Please send the revised photographs on Friday.” Stop. You should see one completed sentence, with no text inserted in another app and no provider API charge.

Now click an empty TextEdit document, press Control–Option–Space, say a different sentence, and press the combination again. You should receive one finished result at the cursor. Check Log to find it again. If it does not paste, use the paste checklist before trying a different recognition engine.

Apple's Voice Control and system Dictation should be off for this first comparison. Otherwise another feature may be typing at the same time.

A private offline setup#

In Providers, choose Local Whisper, select a model, and use Download & Prepare. Wait for readiness before recording. After the download, leave target-language translation and enrichment off to keep both the audio and transcript local. Test once before relying on it away from an internet connection.

Choosing a transcription engine#

The engines share the recording controls, but do not all process audio the same way. Most receive a completed recording; GPT Live streams while you speak. The tables describe Secretary's implementation, not every capability a provider offers elsewhere.

Engine comparison#

These are starting points for a comparison, not an accuracy ranking. The best engine for your microphone, language, and vocabulary is the one that succeeds on your real examples. Cloud model names and availability can change; check Providers in the installed build.

At-a-glance feature matrix#

A separate on-device preview may be shown even when a different engine creates the final answer. It is not a second final result from that provider.

Free recognition does not mean every optional feature is free. Translation sends text to Gemini; enrichment sends text to Gemini or OpenAI. Both can add provider charges after Apple Speech or Local Whisper. See AI Enrichment.

Cost & History displays the app's configured estimates. They are not invoices: current provider prices, free tiers, rounding, taxes, subscriptions, and account-specific rates may differ. Check the provider's billing dashboard before choosing a paid service. No fixed per-minute price in this manual is a purchase promise.

A practical way to choose#

• Start with Apple Speech when speed, privacy, and reliability matter most.
• Use Local Whisper when you need offline multilingual recognition or a shared custom vocabulary without an API charge.
• Try Gemini to compare instruction-guided cleanup and readable formatting.
• Try GPT Transcribe to compare how it handles your names, accents, languages, and technical wording.
• Try GPT Live Transcribe when a streamed cloud preview is useful, after reviewing its estimated cost.
• Treat Meta Muse as a test option until access, billing, language coverage, reliability, and pricing are independently established.

Smart and Verbatim#

Gemini and the two OpenAI engines expose two styles:

• Smart adds punctuation and capitalization, can remove ums and ahs plus obvious abandoned false starts, and can divide longer dictation into readable paragraphs. The three cleanup switches let you turn those behaviors on or off.
• Verbatim asks the provider to preserve spoken wording, including fillers and false starts. Use it for quotations, interviews, legal wording, and testing.

These are instructions, not guarantees. Apple Speech and Local Whisper control most of their own wording and do not offer the same cleanup controls.

Secretary Vocabulary#

Enter one difficult name, product name, acronym, or short phrase per line in Secretary Vocabulary, under Audio & Vocabulary or Test Window. The same saved list is offered only to the currently Live engine:

• Apple Speech receives short contextual hints.
• Local Whisper receives private on-device prompt tokens.
• Gemini and GPT Transcribe receive prompt guidance.
• GPT Live receives live keyword hints and prompt guidance for its fallback.
• Meta receives keyword metadata.

The list does not modify Apple Dictation or Voice Control vocabulary. Recognition hints improve the odds but no engine guarantees a spelling. Use Import CSV and Export CSV to move lists between installations.

Give a difficult name a fair test#

Try a short list such as Supabase, SwiftUI, TinyCal, and iWatermark+. Say an ordinary sentence containing the names, then compare the same words alone. A term can be easier in context than as an isolated sound. Speak naturally at a comfortable microphone distance; stretching every syllable can change the sound the recognizer expects.

Keep only genuinely useful terms in the list. A first transcription can take longer while a local model prepares; faster subsequent runs do not prove that the app learned your voice. Repeating a word is not currently a correction command and does not cycle through homophones. Correct the document normally, and update the vocabulary list when a stable spelling matters.

Hotkeys#

The hotkey menu uses behavior names so it is clear whether to press, hold, or double-tap.

Assigning or registering a combination does not prove macOS will deliver it globally. Another app or a macOS feature may intercept it without identifying itself. Test Window separates four possible problems: assignment, registration, delivery, and actual recording/transcription.

The Secretary window and menu#

Menu bar menu#

Click the Secretary icon in the macOS menu bar.

1. Start Recording changes to Stop & Transcribe while recording and to Transcribing… while waiting for the result.
2. Open Test Window opens Preferences directly to the safe testing workspace.
3. Cancel Active Transcription appears while a provider request is running. It stops that job; it is not a way to recover words already pasted into another app.
4. Preferences… returns to the saved preferences page.
5. Purchase… opens purchasing information.
6. Check for Updates… checks for a newer release when updates are available.
7. Quit The Secretary releases the microphone, hotkey, and Local Whisper memory, then quits.

The older Open The Secretary, License…, and About The Secretary menu entries have been removed. License and About are still available in Preferences.

Main window#

• General shows Ready, Recording, or Transcribing status; the current or completed text; a copy button; and a microphone button.
• Log searches saved transcription, translation, and enrichment text and can copy previous results.
• The gear opens Preferences.
• Cancel stops an active provider request.

The main window is useful for watching and copying results, but everyday dictation is designed around the global hotkey and the field where you were already working.

Know when recording has stopped#

Look for Recording or listening feedback when capture begins, Transcribing… after stopping, and Ready when the job finishes. A green LIVE engine badge alone does not confirm that the microphone is open. A live preview alone does not identify which application is typing into your document.

With a press-to-start/stop shortcut, stopping requires the second press. Sending a message, changing fields, or falling silent does not replace it. With Hold to Talk, release the configured key. Always check the state before leaving a recording unattended.

Preferences#

Preferences has a searchable sidebar. Purchase and License remain anchored at the bottom while the other pages run sequentially from the top. The green LIVE badge at the upper right appears on every page and names the active engine. Click it to return to Providers. Changes apply immediately.

The Ticker button rolls Preferences into a compact title-bar status display; double-click the ticker to restore the full window. The sun or moon button changes the app's appearance.

General#

• Startup & Updates: Optionally open Secretary at login and manually check for updates.
• Speech Recognition Hotkey: Select a preset or record a custom combination. The instruction underneath describes the complete gesture.
• Menu Bar Icon — Beta Experiment: Compare Microphone and The Dictator at their real menu-bar size. Microphone is the default. macOS automatically renders the selected icon black in Light Mode and white in Dark Mode; green means listening and orange means transcribing.
• App Icon — Beta Experiment: Compare the enlarged and classic Dictaphones with Typewriter, Secretary, and Steno. The choice changes the running app in the Dock and app switcher and persists across launches. Finder continues to display the icon embedded in the application bundle.
• When Secretary finishes: Choose Paste final (the default), Don't paste, or Preview + paste. See the delivery table below. This setting does not turn Apple's Dictation or Voice Control on or off.
• Gremlin voice: An optional on-device system voice for future guidance. Test Voice speaks locally and sends no text to a provider.
• Actions & Feedback: Control automatic paste into the current cursor location in a text field or text area, whether the finished transcription remains ready to paste again from the clipboard, and recording-state sounds. If Keep finished transcription on clipboard is off, Secretary temporarily replaces the clipboard to send Command-V and then restores the previous image or text. It never empties the clipboard when recording begins.

Command–Comma and Preferences… use the existing preferences window and return to the last selected page. Open Test Window deliberately opens the testing page instead.

Secretary's preview is displayed in its own interface. It is not live insertion into another application's field. If Apple Voice Control is already typing there, Paste final and Preview + paste can add a second block at the end. Don't paste avoids that second Secretary insertion while retaining its result. Turn Voice Control off for a clean engine comparison.

The master Automatically paste text… switch can disable the final paste in every mode. Test Window and Watch Folder do not paste into the active app regardless of this switch.

Providers#

• Transcription Engine: Click an entire provider row to select it and reveal its strengths, tradeoffs, languages, translation behavior, privacy, and estimated cost. When the shield and selected line are both green, the provider is ready to use. A separate shield column reports readiness: green means the local engine is ready or a cloud credential is saved; orange means permission, experimental access, or account verification still needs attention. A saved cloud key is not proof that provider billing or access will succeed, so verify it in Test Window.
• API key: Gemini, Meta, and OpenAI keys are stored in Apple Keychain. Revealing or changing a key requires Touch ID or the Mac password, and the editor locks again after 60 seconds.
• Meta speaker diarization: Optionally labels different speakers when the endpoint supports the request.
• Local Whisper model: Choose Fast — Base Multilingual or More Accurate — Small Multilingual, then download and prepare it once. Choose spoken-language output or local speech-to-English translation. Sound-description captions are hidden by default. Reduce text invented during silence rejects brief steady-noise recordings and trims safe outer silence without deleting words after a thinking pause.
• Advanced Whisper Settings: Normally leave these collapsed. You can lock a spoken language, adjust temperature and warmer retries, change silence and decoder-confidence thresholds, enable experimental pause-based VAD splitting, or add window-end clipping. Reset to Defaults, beside the section title, returns to automatic language, temperature 0, WhisperKit confidence defaults, no VAD splitting, and no clipping. The Secretary automatically supplies 16 kHz mono audio. Beam size is not shown because the installed WhisperKit decoder does not expose beam search through this path.
• Optional Translation: Select target languages for a separate Gemini text-translation request after transcription. Open the Transcriptions folder from this section.
• Formatting Style: Gemini and OpenAI offer Smart or Verbatim. Smart exposes punctuation, filler-removal, and paragraph switches.
• Phonetic & Phonemic (IPA): Gemini can add phonemic, phonetic, or both IPA representations beneath the transcript.

Audio & Vocabulary#

• Choose the microphone used by every engine. Switching providers does not change it.
• Start Meter temporarily opens the selected microphone and shows its level. Gray is too quiet, green is useful, and red indicates clipping. Stop Meter, start a recording, or leave the page to release it.
• Read any microphone error shown beneath the meter.
• Edit Secretary Vocabulary one phrase per line, or import/export CSV files.
• Open Recoverable Recordings… shows audio retained after incomplete or failed attempts. Delete All… asks for confirmation before permanently deleting recoverable WAV files; see Where are my files?.

The meter is not an always-on recording. It saves and sends no audio, and Secretary releases the microphone whenever recording and the explicitly started meter are both idle.

Speech Health#

Speech Health reports what Secretary can verify rather than guessing:

• Microphone authorization, selected input, current capture state, sample rate, channel count, last audio event or error, and recovery count.
• Apple Speech availability and on-device recognition support.
• Local Whisper installation and memory status, with the optional Release Whisper Memory Now control. It unloads the model from RAM without deleting its downloaded files. The Secretary also releases it automatically after ten idle minutes or when macOS reports memory pressure; the next Local Whisper transcription reloads it.
• Hotkey registration and Accessibility status.
• Apple settings links open macOS Microphone Permission, Dictation, and Voice Control panels. Those settings belong to Apple, not Secretary.
• Refresh Status re-reads the diagnostic display. It does not reset preferences, download a model, repair recognition, or clear History.
• Voice Control guidance and Export Vocabulary for Voice Control… for a list you can import yourself. No automatic synchronization is performed.

Cost & History#

• Shows total recordings, duration, word count, session cost, current-month estimates, all-time estimates, and provider-by-provider totals.
• Searches transcript, translation, and enrichment text.
• Filters the history by provider.
• Opens the Transcriptions folder, copies a result, shows a sanitized API/cost inspector, or clears all history after confirmation.
• Estimates exclude taxes, free tiers, negotiated rates, and provider-side rounding.
• Dictation Length Science summarizes measured completed recordings, with the limits described in Recording length.
• Storage & Privacy controls optional plaintext exports and opens the separate recoverable-audio folder.

Log#

Log is a larger second entrance to the same transcription history found in Cost & History. Search words from the transcript, translation, or enrichment, or select an engine to narrow the list. Choose All when an older entry is missing from a provider filter; a live session may have completed through a file-transcription fallback.

Use an entry's Copy Entry control to copy that dictation without selecting the entire log. The copied entry also includes available translation and enrichment. Enrichment appears beneath the matching entry rather than in the text originally pasted into your document. Inspect the entry's status if analysis is still pending or failed.

The clear-history confirmation distinguishes Clear History Only from Clear History and Exported Text Files. Both are destructive and cannot be undone inside Secretary. Neither deletes recoverable audio. Cancel if you still need the evidence for a support request.

AI Enrichment#

Enable any combination of Summary, Action Items, Key Points, Chapters, and Sentiment. Enrichment uses one additional Gemini or OpenAI text request after the original transcription is delivered. Automatic provider selection prefers OpenAI after an OpenAI transcription, otherwise Gemini when available, then OpenAI.

When Generate enrichment automatically is on, the selected analyses run after every successful transcription, regardless of which speech engine produced it. The completed transcript text is sent to the selected cloud enrichment provider; the audio recording is not sent for enrichment. The analysis does not alter or append itself to the pasted text. Open Log to see the enrichment beneath its matching transcription, including pending or failed status and a retry button when needed.

Enrichment is off by default, never delays the initial paste, and creates a separate token-based provider charge that is not included in the audio-duration estimate. The app warns before enabling it and shows a per-entry estimate calculated from the provider's reported token counts. When Apple Speech or Local Whisper performed the free transcription, the extra enrichment estimate is shown in red so the separate cloud charge is hard to miss. Provider invoices, free tiers, taxes, and future price changes remain authoritative.

Automatic chooses the enrichment company. Generate enrichment automatically decides whether to request it after each completed transcription. These are different controls. Apple Speech and Local Whisper can both receive enrichment because it analyzes their finished text, not because their speech models have built-in summarization.

To try it once, save a Gemini or OpenAI key, enable automatic enrichment and Summary, then dictate a short plan in Test Window. For example: “Send the photographs on Friday. Ask Maya which three she wants printed.” Open the matching Log entry after the transcription finishes. A useful analysis would summarize the plan; with Action Items also selected it may list the two tasks. These are illustrative results, not guaranteed wording. Turn automatic enrichment off again if you do not want further text sent or charged.

Sentiment is an interpretation of the transcript's wording, not a reliable measurement of your emotion or vocal tone. Pretending to sound angry will not necessarily change neutral written words. Check the analysis in Log, not the pasted sentence.

Watch Folder#

• Choose a folder and click Start Watching.
• Secretary waits for copying to finish and queues stable AAC, AIFF, FLAC, M4A, MP3, MP4, MOV, and WAV files.
• Files use the currently Live engine and normal history, cost, translation, export, enrichment, and clipboard settings.
• Watch Folder never pastes into the active application.
• Move source files to Trash after successful transcription is optional and off by default. Failed files are kept; trashed files remain recoverable until Trash is emptied.

Test Window#

Test Window puts the controls needed for fair comparisons in one place:

• Switch providers and see their detailed capabilities.
• Edit Secretary Vocabulary.
• Change Smart/Verbatim, Apple punctuation, Meta speakers, Whisper output, Whisper sound descriptions, and Whisper model settings.
• Record with the large microphone button without pasting into another app.
• Select or create a hotkey, arm the global test, and see whether macOS actually delivers it.
• Keep the transcription result and log inside Test Window until you end the hotkey test.

For normal work, end the test and use the configured hotkey in the destination field.

About#

About shows the installed application icon, product name, version and build number, copyright, acknowledgements, and direct links to the product website, support, privacy information, user manual, and Sparkle update check. Join the Newsletter opens Plum Amazing's signup form with a source tag identifying the app; subscription remains an explicit opt-in on the website.

Defaults and a careful reset#

Leave Advanced Whisper Settings alone until a repeatable test identifies a reason to change one. Its Reset to Defaults restores only those advanced decoding settings; it does not reset the whole app, remove keys, clear vocabulary, or delete recordings.

For a clean everyday baseline, select Apple Speech, System Default Microphone, Control–Option–Space, and Paste final. Enable automatic paste and clipboard copy. Leave translation and automatic enrichment off. For Local Whisper, prepare the selected model, keep spoken-language output, keep silence protection on, and leave experimental chunking and clipping at their defaults. This is a checklist you apply deliberately, not a promise that one reset button changes every setting.

Change one setting at a time and repeat the same test. Keep your existing vocabulary and any recovery audio; they may be the evidence that explains the problem.

Translation, enrichment, and automation#

Target-language translation#

Transcription asks what was spoken; translation asks how that text should read in another language. They are separate operations.

1. Save a Gemini API key.
2. Select target languages under Optional Translation (Powered by Gemini).
3. Use any engine for the original transcription.
4. Secretary sends the finished text—not a second copy of the audio—to Gemini. Review the translation in History. Separate text files are written only when plaintext export is enabled.

Apple Speech and Local Whisper can keep the original audio on the Mac, but enabling target-language translation sends the completed text to Google and adds time and API cost.

Local Whisper speech-to-English#

Local Whisper has a separate Translate Speech to English output mode. That conversion is performed locally from supported non-English speech. It is different from Gemini target-language translation.

AI enrichment#

Enrichment starts after successful transcription, including workflows that intentionally do not paste. Results are attached to the matching Log/history item rather than pasted into the destination document. Review generated summaries, chapters, sentiment, and action items before relying on them; the service can misunderstand or omit details.

Watch Folder#

Watch Folder is intended for existing media and unattended batches. Keep source files until you have checked the resulting transcript. The optional Trash setting is recoverable, but it should be enabled only after the workflow has been tested with copies.

For a first batch, choose a new folder containing copies of two short audio files. Select the desired engine and leave automatic trashing off. Start watching, then wait for both History entries. Check their text and costs before adding more files. Stop watching before changing your batch workflow. A recording failure should leave its source available for another attempt.

Apple Voice Control: a companion, not a hidden engine#

Voice Control is an Apple accessibility feature with its own recognition, vocabulary, and spoken commands. Secretary does not take over its engine or automatically start, stop, or put it to sleep. The two applications may listen at the same time.

You can create a Voice Control command that presses Secretary's current start/stop hotkey. With Secretary set to Control–Option–Space, open macOS Voice Control's Commands…, create a phrase such as “Secretary start or stop,” and assign the action that presses that combination. Labels vary by macOS version. Test the command while watching Secretary's recording state. The phrase is a toggle: the same command starts an idle recording and stops an active one. It is not two state-aware Start and Stop commands.

Apple's wake/sleep commands affect Voice Control, not necessarily Secretary. Stop Secretary explicitly before leaving. Spoken commands such as “select” may edit the current document rather than become dictated words; that behavior belongs to Voice Control.

Use Don't paste when you want Voice Control to type into the field while Secretary keeps its separate transcript. For a fair Secretary-only test, turn Voice Control and system Dictation off. The Fn key can trigger Apple's Dictation independently, so Control–Option–Space is the safer initial comparison.

Export Vocabulary for Voice Control… saves a list for manual import into Apple's vocabulary controls. It does not transfer voice training or keep the lists synchronized. You choose whether to import it; your existing Apple vocabulary remains separate.

Shortcuts on Mac#

The current actions are Start Recording, Transcribe Audio File, Get Usage and Cost Summary, and Translate Text with Gemini. Their availability depends on the installed app and system version.

• Start Recording currently opens Secretary; it does not reliably begin capture by itself. Start with the microphone or configured hotkey and verify the recording state. Do not use this action as an unattended recorder.
• Transcribe Audio File takes a file and returns text. In the present implementation Apple Speech uses Apple recognition; GPT and GPT Live selections use OpenAI file transcription. The other selections use the Gemini path. It is not equivalent to the Mac Watch Folder, and selecting Local Whisper here is not a guarantee of local processing. Use Watch Folder with Local Whisper when that privacy boundary matters.
• Get Usage and Cost Summary returns recorded counts, words, duration, and estimated cost—not a provider invoice.
• Translate Text with Gemini takes text and target language names and needs a Gemini key. Check that the returned language is correct before automating delivery; a failed translation can return the original text.

For a small, low-risk Shortcut, add Get Usage and Cost Summary, then Apple's Show Result action. Run it and compare the displayed totals with Cost & History. This reads local usage; it does not send a support report or buy anything. File and translation actions may incur charges, so test them with non-sensitive material and inspect the output before adding a sharing action.

Purchase, licensing, and provider charges#

Current beta status#

The Mac beta includes separate Purchase and License panels in Settings. Until Plum Amazing adds the final price, checkout URL, and approved Lemon Squeezy product IDs to the signed release, the Purchase panel says Beta access is free and the application stays unlocked. This prevents unfinished store configuration from blocking beta testing.

This beta has no purchase reminders or evaluation countdown. Evaluation days begin only in a later commercially configured build with a working checkout. Optional cloud transcription and enrichment can still incur charges from your provider account.

The shareware policy prepared for that later build keeps all features usable: after 30 full evaluation days, quitting shows a purchase invitation; after 45 days, recordings longer than five minutes also receive a separate invitation after the text has been delivered. Purchase License is available immediately. Not Yet counts down five seconds, then dismisses the invitation (and allows quitting when the invitation was shown on quit). Choosing Purchase opens checkout and leaves the app running. After 60 days the wording becomes more direct, but the app does not delete API keys, stop saving preferences, or remove recordings or transcripts. Activated customers do not see these invitations.

When offered, the optional Student / community discount uses a separate 20%-off checkout. Students and anyone who needs a little help may choose it; student verification is not required. There is no active checkout or discount link in this beta. A separate checkout link is not private and can be shared.

When commercial licensing is configured, enter the key from the Lemon Squeezy purchase email in License. The app accepts only the configured Secretary product IDs and stores the key and Lemon Squeezy instance ID in Keychain. Verify License checks the saved activation. A network outage keeps the saved local activation instead of revoking it.

The Version 1 offer described in the Purchase panel is a permanent 1.x purchase for up to two Macs, including 1.x updates. Check the actual checkout terms when purchasing becomes available; this beta is not a completed sale or a promise of a final price. Future major versions, add-ons, hosted services, and separate products are not automatically included.

Deactivate This Mac releases this computer's activation at Lemon Squeezy. Local license data is cleared only after the store confirms deactivation. Deactivating a Mac does not invalidate or refund the permanent Version 1 purchase; it frees an activation so the same key can be used on another Mac.

Provider accounts are separate#

Buying or licensing Secretary will not include Google, Meta, or OpenAI usage. Cloud engines use the customer's own provider account and API key, and the provider bills that account directly. Meta may require both API billing and a separate subscription attached to the credential. Apple Speech and Local Whisper have no per-minute provider charge.

Secretary stores cloud credentials in Apple Keychain. The app can use a saved key while its editor is locked; revealing or changing it requires device-owner authentication.

Updates#

The Mac app includes Check for Updates… and Automatically check for updates. Updates are not installed without your approval. If this beta says updates are not yet published, a manual check cannot manufacture a newer version; use the next build supplied by Plum Amazing.

Recover a purchase or move to another Mac#

Keep the purchase email and key. If you cannot find them, contact Plum Amazing Support using the purchase email address. Never post your key or provider credentials in a public issue.

When licensing is enabled, use Deactivate This Mac on the old computer and wait for confirmation, then activate the new Mac with the same key. Deactivation releases a device seat; it is not a cancellation or refund. If the old Mac is unavailable or the activation limit is reached, contact support instead of repeatedly deleting local settings. A provider API key is unrelated to the Secretary purchase key.

iPhone and iPad#

Main app#

• Tap the microphone to start and stop recording.
• Use the settings sheet to choose an engine, target languages, and API credentials.
• Live text is hidden by default, as on the Mac.

Secretary keyboard#

The current beta may still list this extension as Transcriber Keyboard.

1. Open Settings → General → Keyboard → Keyboards.
2. Tap Add New Keyboard… and choose the Secretary/Transcriber keyboard.
3. Open the keyboard's settings and review Allow Full Access when required for shared settings and cloud access. Full Access is not a promise that iOS will allow microphone capture in a keyboard extension.
4. In a text field, use the Globe key to select the keyboard. Treat its recording workflow as experimental in this beta; if capture is unavailable, dictate in the main app and copy the result instead.

Share extension#

Share a supported audio file from Voice Memos or Files to the Secretary share extension. The current beta may call it Transcriber Share. The extension transcribes the file, writes the normal logs and exports, and copies the completed output.

Shortcuts and Action Button#

The shared Shortcuts actions are described in Shortcuts on Mac, including their current limitations. An Action Button can launch a chosen Shortcut on supported iPhones, but assigning Start Recording does not yet guarantee that audio capture begins. Confirm the app is actually recording. Mac-only Local Whisper, Watch Folder, hotkey, and enrichment instructions do not imply feature parity in these mobile extensions.

Privacy and stored information#

• Apple Speech and Local Whisper process audio on the device when their local requirements are satisfied.
• Gemini, Meta, GPT Transcribe, and GPT Live send audio to their respective providers.
• Target-language translation sends completed transcript text to Gemini.
• AI enrichment sends completed transcript text to Gemini or OpenAI.
• API keys are stored in Apple Keychain, not in preferences text files.
• History and usage estimates are stored locally in app preferences; transcript and translation exports use the current beta's Transcriptions folder.
• The microphone meter does not save or transmit its audio.
• Watch Folder security access is limited to the folder selected by the user.

History, exported text, and retained WAV files can contain sensitive information. They are not made private merely by selecting a local engine: anyone with access to your Mac account or backup may have access to local files. Decide what to keep and remove it deliberately. Plaintext exports are ordinary readable files, not encrypted documents. Deleting an item in Secretary does not remove copies you made elsewhere or a provider's records.

Switching to a local engine stops using cloud recognition for subsequent dictations; it does not automatically disable optional translation or enrichment. Review both settings when you want an entirely local workflow. For cloud retention and account controls, consult the service's current policy and your own account settings.

Before recording another person, obtain any consent required by local law, workplace policy, or the service being used.

Troubleshooting#

A hotkey does nothing#

1. Open Test Window and select the same hotkey.
2. Read the full gesture instructions.
3. Click Arm Hotkey Test and try it.
4. If no event arrives, grant Accessibility permission for modifier-only shortcuts, check macOS Keyboard settings for Fn behavior, or choose another combination.
5. If an event arrives but recording does not begin, open Speech Health and check microphone permission and the selected input.
6. If recording works but no text returns, troubleshoot the selected engine rather than the hotkey.

Secretary can detect registration and received events, but macOS does not provide a complete reliable list of every app or operating-system feature that may intercept a shortcut.

Recording works but text is not pasted#

• End any armed Test Window hotkey session; Test Window deliberately keeps results inside its panel.
• Turn on Automatically Paste text at current cursor.
• Select Paste final or Preview + paste; Don't paste deliberately saves without inserting a finished block.
• Put the insertion point in an editable field before starting.
• Check Accessibility permission if paste simulation is blocked.
• The completed result should still be available in the clipboard, main window, and Log when their corresponding settings are enabled.

Duplicate text or duplicate menu icons#

Quit extra development, beta, or release copies of Secretary/Transcriber. Only one process should own the global hotkey. Then turn off Apple's Voice Control and system Dictation for a controlled test. Either can insert text independently while Secretary records, even if Secretary's preview is hidden.

If you intentionally use Voice Control for live typing, set Secretary to Don't paste to avoid adding its finished transcript to the same field. If you want only Secretary's final result, keep Paste final, disable the other live typist, and use Control–Option–Space rather than a shared Fn gesture. Changing Secretary's setting cannot stop a separate Apple feature.

An old clipboard image keeps returning#

When Keep finished transcription on clipboard is off, Secretary preserves and restores the item that was on the clipboard before it pasted. That is why an image can reappear. Turn this preference on when you want each finished transcript to replace the image and remain on the clipboard. Secretary does not clear the clipboard at hotkey-down because a failed or cancelled transcription would otherwise destroy something you copied.

Apple Speech returns only part of a recording#

• Confirm on-device recognition is available in Speech Health.
• Speak clearly and test without competing audio.
• Compare the same recording with Local Whisper or a cloud engine.
• Apple controls recognition segmentation and may revise interim text. Secretary preserves the fullest accumulated result it receives, but it cannot reconstruct speech Apple never returned.

A cloud engine is slow#

Cloud file engines begin their provider request after the key is released. The first request may be slower because of connection setup, file upload, model warm-up, or provider load. GPT Live is the only cloud engine in this build that streams its own transcript while you speak. Apple Speech is normally the lowest-latency option because it runs locally.

Meta returns HTTP 402 or billing_not_configured#

Meta rejected the request because of billing or account access. Verify billing and any required plan subscription for the exact account associated with the key, then test again. That error alone does not prove transcription succeeded or validate every audio setting. Secretary cannot override Meta's account checks.

Local Whisper says a resource path does not exist#

Open Providers, select Local Whisper, and run Download & Prepare for the selected model. If it is marked ready but fails, use Speech Health to release Whisper memory, reselect the model, and prepare it again.

Whisper inserts [bell], [music], or similar text#

Turn off Include descriptions of sounds. The Secretary removes common standalone sound captions—including Whisper's [BLANK_AUDIO] marker—while preserving ordinary dictated bracketed text.

Whisper adds an unrelated phrase, credit, or URL#

Keep Reduce text invented during silence turned on. Whisper can generate plausible text during silence or non-speech; this does not mean another person's transcript was received. Secretary trims safe outer silence and stops very short recordings containing only steady fan or room noise before decoding. It does not blindly delete URLs, because a person may genuinely dictate one, and it does not remove words merely because they follow a thinking pause. The Test Window should instead say No speech detected. Brief background noise was ignored. If invented text still appears after real speech, retain the recording and report the exact added text so the thresholds can be tested without guessing.

The microphone level is blank#

Open Audio & Vocabulary and click Start Meter. Secretary intentionally does not keep the microphone open while idle. If the meter fails, choose System Default Microphone, verify Microphone permission, reconnect the device, and read the error shown in Speech Health.

Where are my files?#

Completed text remains in History inside Secretary. Additional plaintext files are off by default; enable Export completed transcriptions as text files under Cost & History → Storage & Privacy when you want copies in ~/Documents/Transcriber/Transcriptions/.

Audio from a successful transcription is removed. Audio from an incomplete or failed transcription is retained under Recoverable Recordings so it is not silently lost. Audio & Vocabulary and Cost & History show its file count and size and provide separate Open and confirmed Delete controls. Clearing History never deletes recoverable audio.

To recover a failed attempt, choose Open Recoverable Recordings…, sort by date, and play the likely WAV to verify it is yours. File names alone do not prove which recording you want. Make a copy before further testing. You can then place that copy in a watched folder and try a different engine. A retained file is not a guarantee that the microphone captured every word; listen to the beginning, middle, and end.

When you no longer need the evidence, use Delete All… beside the recovery controls and confirm the permanent deletion. This removes recoverable WAV audio, not History, exported text, downloaded Whisper models, or keys. Do not delete the entire application-support folder. Deletion is unavailable during an active recording; stop first. Keep a backup of anything important.

Creating a support report#

Open Speech Health → Support and choose Generate Support Report. Secretary shows the complete report before anything leaves the Mac. You may copy it, save it as a text file, or open Plum Amazing Support.

The default report contains app and macOS versions, architecture, permission states, selected provider, relevant settings, credential-presence booleans, non-content recording measurements, warnings, and recent provider/duration diagnostics. It excludes transcript and enrichment text, raw hypotheses, audio, API keys, clipboard contents, usernames, hostnames, and file paths. Nothing is emailed or uploaded automatically.

Send the reviewed report through Plum Amazing Support when asked. Add what you expected, what happened, the engine/model, whether Voice Control was on, and whether the problem repeats. A short example with permission to inspect it is more useful than changing many settings at once. Audio and transcripts are separate attachments you choose to share; never assume the report already includes them.

Enrichment is missing#

Check Generate enrichment automatically, at least one selected analysis, and a usable Gemini or OpenAI key. Automatic provider selection does not enable the feature by itself. Open the matching Log entry and wait for its pending request; review a failed status and retry only when you want another request. Apple Speech and Local Whisper are supported because enrichment operates on their text. Nothing is appended to the original pasted dictation.

Whisper misses the beginning or a paragraph#

Keep the recoverable audio and compare it with the transcript. If the words are absent from the WAV, inspect the selected microphone and capture state. If they are audible but missing from the text, record the model and any diagnostics, restore only Advanced Whisper Settings to their defaults, and compare the same saved file through another engine. If Log has the full text but the document does not, inspect paste mode, the target field, and competing dictation tools.

Do not diagnose all three cases as a duration limit. Use the three-marker test and provide the evidence with a support report. A silence-protection setting should not be an excuse to lose real speech after a thinking pause.

Memory use and the first transcription#

The app's memory total varies with the engine, downloaded model, recording length, and macOS caches. A single observation such as 310 MB is not a fixed budget for every user. Local Whisper can keep its model in RAM for reuse and releases it after ten idle minutes or memory pressure. Release Whisper Memory Now unloads it without deleting the download; the next use must load it again and may take longer. macOS may not immediately show an identical drop in Activity Monitor because framework and system caches are accounted for separately.

Before concluding that memory is leaking, compare the same engine and short task over repeated runs, then let the app idle or release the model. Report sustained growth with the app/build, model, and task—not just one screenshot. Faster warm runs are expected; improved accuracy on a second reading is not proof of learned pronunciation.

Reference links#

• Google Gemini audio transcription
• OpenAI speech-to-text guide
• Apple Speech supported locales
• Plum Amazing Support
• Apple on-device speech recognition
• OpenAI Whisper model card

Provider capabilities and prices change. Provider documentation and billing dashboards are authoritative; Secretary's descriptions and usage totals explain this build and provide estimates for comparison.

Glossary#

Revision history#