# Changelog
Source: https://docs.stenox.app/changelog
The main features and changes in each Stenox release.
## 2.2.1
### Enhancements
* **Easier profile editing.** Collapsible sections keep transcription, enhancement, and auto-activation settings organized, with more compact Polish controls.
* **Recognizable app profiles.** App-specific profiles and activation rules now display application names and icons.
* **Safer text insertion.** Improved handling of focus changes when inserting dictation or replacing text with Polish.
### Bug fixes
* Fixed Polish copying its result instead of updating the dictated passage in some supported text fields. Undo can now restore the original dictation in those fields.
* Fixed previous clipboard contents not being restored after some successful Polish replacements.
## 2.2.0
### New features
* **Online research with CLI tools.** When you use Codex, Claude Code, or Antigravity for meeting AI, Ask Stenox and generated insights can search online, with source links and clearly labeled web context. Online search is available only through these CLI tools; local models do not browse the web.
* **Polish with your subscription.** Choose Codex, Claude Code, or Antigravity for optional dictation polishing, with separate model and reasoning preferences from meeting insights.
### Enhancements
* **Exclude unwanted speech from insight generation.** Exclude a speaker or an individual passage from insights, answers, and AI notes while keeping the transcript and recording. Restore exclusions at any time.
* **Faster local search.** Improved indexing and retrieval reduce delays and timeouts when searching larger collections of dictations and meeting notes.
* **More flexible title matching.** Find saved meetings with variations in spacing or punctuation, with typo recovery for otherwise empty keyword searches.
* **Clearer app layouts.** Refined Settings, Vocabulary, and activity views, with more compact saved-meeting rows and search controls.
* **Freshened up the logo.**
### Bug fixes
* Fixed long meeting titles and toolbar controls being clipped when Ask Stenox is open.
* Fixed Polish model selections being replaced when switching profiles or accounts, or switching between local and subscription providers.
## 2.1.0
### New features
* **Search your saved work.** Find dictations and meeting notes by words, phrases, or meaning. Filter by date, app, or content type and open the matching passage. Search runs locally, with an optional multilingual model.
* **Remember vocabulary corrections.** Save corrections to names and unfamiliar terms directly from supported text fields. Apply them to one app or across all apps, and manage them in Vocabulary.
* **Replay meeting recordings.** Keep audio locally for new Notetaker sessions, jump to transcript timestamps, adjust playback speed, and follow the conversation. Remove recordings without deleting transcripts or notes.
### Enhancements
* **Easier meeting-note editing.** Edit session notes directly, review updated AI drafts while preserving your edits, and export personal notes as Markdown.
* **Better Notetaker navigation.** Read earlier passages without being pulled back to the latest speech. Saved answers now show confirmation, and failed questions can be retried.
* **More compact dictation controls.** A smaller completion popup keeps Polish, Copy, and Undo within reach.
* **Clearer license settings.** See how many Macs are active on your license, with simpler activation and removal controls.
### Bug fixes
* Fixed file transcription being interrupted when switching pages, preserving selected files, settings, and results.
* Fixed question drafts being lost when navigating away from Ask.
* Fixed duplicate pasting in Chrome and insertion issues caused by focus changes.
* Fixed canceled insertion operations unexpectedly restoring focus.
* Corrected completion-popup dismissal timing and Polish controls that did not reflect current settings.
* Fixed History losing the distinction between original recognition and cleaned text in some dictation paths.
* Improved recovery of buffered meeting speech after capture interruptions and handling of unsaved work before quitting.
* Fixed permission checks disrupting recording shortcuts and cancellation.
## 2.0
* **Meetings with Notetaker.** Capture microphone and system audio, follow speaker-labeled conversations, and keep insights, personal notes, references, and answers together. Preparation questions guide insights, which you can also generate on demand. [Explore Notetaker](/meetings/overview).
* **Audio and video transcription.** Turn files into transcripts, with speaker labels and timed exports where the selected provider supports them. [Transcribe a file](/setup/file-transcription).
* **Dictation first, Polish when you choose.** Get your words into an app, then request an optional rewrite. Copy and Undo help you review the result; filler removal has its own setting. [Try dictation](/first-dictation).
* **Local and cloud choices.** Choose on-device speech and AI models, your own supported provider keys, or a connected subscription for meeting insights. Speech and insight processing have separate selections. [Choose providers](/providers/overview).
* **Updated setup and recovery.** Onboarding prepares your selected local speech model. Recover access to saved provider and licensing credentials from Settings. [Installation](/installation) · [Access recovery](/guides/upgrading#recover-provider-and-license-access).
* **Trial and license access.** Try the full app for seven days. Valid earlier paid licenses cover Stenox without another purchase and retain their original rights and device allowance. Existing free 1.x installations remain usable. [Release and access](/pricing/overview) · [Upgrade guide](/guides/upgrading).
# Your first dictation
Source: https://docs.stenox.app/first-dictation
Record and insert speech, then choose whether to polish it.
## Before you begin
Complete [installation](/installation), enable microphone and Accessibility permissions, and wait for the selected speech model to be ready. Onboarding automatically prepares the local speech selection; the new-install default is **Parakeet v2 (English)**. Profiles and saved settings can change it.
Start the [seven-day trial or activate your license](/pricing/overview) before starting new work.
Open Notes, TextEdit, your browser, or another app and click an editable field.
Hold **Fn/Globe**, or press **Option–Space** to toggle recording. Use your saved shortcut if you changed it in **Settings → General**.
Say a short sentence. Release the hold shortcut, or press the toggle shortcut again. Stenox transcribes the audio and applies enabled filler removal and vocabulary replacements.
Text is inserted at the target or copied according to your output setting. The completion controls offer **Copy**, **Polish**, and **Undo** when available.
## Polish when you want to
Open **Text cleanup** and enable **Polish on demand** to choose a local or Gemini model. After dictation, click **Polish** to request a rewrite using that selection. With polishing disabled, the control leads you to its settings.
**Remove filler words** is independent of AI polishing. It removes clear hesitations without loading a language model. Dictation delivers this deterministic result first; it does not automatically run the selected AI model.
The default local cleanup selection is **Gemma 4 E4B**, with polishing initially off. [Local AI cleanup](/providers/llm/mlx-local) lists the available choices and their limits. Download and prepare a local model before using Polish; the Polish operation does not silently download a missing model or switch to a cloud provider.
## Review and recover
History retains the original dictation. When Stenox can verify the inserted text, **Undo** can remove a dictation or restore its pre-polish text. If the target or clipboard changed, Stenox may offer the polished text for **Copy** instead of replacing it. Review names, numbers, and important details before using an AI rewrite.
Dictation and meeting capture cannot run at the same time. End the active meeting before starting a dictation.
Next: [shortcuts](/setup/hotkeys), [profiles](/profiles/overview), or [word replacements](/setup/vocabulary).
# Set up cloud processing
Source: https://docs.stenox.app/guides/cloud-setup
Choose cloud features intentionally and verify a short sample.
1. Decide which job you want to run remotely: speech, optional dictation Polish, or meeting insights.
2. For speech or Gemini Polish, add the provider key in **Settings → API Keys** and select it in the relevant profile or Text cleanup controls.
3. For subscription meeting insights, connect the provider in **Settings → Notetaker** and choose a runtime-supported model and reasoning setting.
4. Try a short sample and check output, account access, quota, and usage before a longer session.
Cloud speech sends audio; cloud cleanup and insights send required text and context. Provider usage is separate from Stenox app access. Do not rely on historical free-credit or pricing figures from older guides.
A local speech + cloud insight setup still sends meeting text. Insights and refinement also include an excerpt of Personal notes and preparation questions. Google-linked sessions use the selected insights provider, which can receive included calendar/reference context and derived information. [Each operation uses different inputs](/meetings/processing#context-and-failures).
Use the session's **Insights model** menu to change providers or models during capture. **Generate insights** requests a review even with **Automatic insights** off. Switching preserves recording and existing insights, but does not sign out providers or recall earlier transfers.
If a service is unavailable, the app does not silently move the job to another account. Notetaker preserves capture and saved text when optional insight generation fails; inspect the error and explicitly retry or change the provider.
See [supported BYOK services](/providers/cloud-byok), [meeting providers](/meetings/processing), and [troubleshooting](/guides/troubleshooting).
# Set up local processing
Source: https://docs.stenox.app/guides/privacy-first-setup
Prepare local speech, optional cleanup, and meeting insights for use on your Mac.
In your active profile, select Parakeet, Whisper, or Cohere Transcribe and a supported language. Wait for the model to download and load.
Leave Polish off if you only need transcription, filler removal, and replacements. If you want Polish, select **Local** in Text cleanup and prepare its model.
Choose **On-device** insights in Settings → Notetaker or the session's **Insights model** menu. Allow the separate insight and speaker models to prepare before relying on them. Turning automatic insights off is separate: **Generate insights** still uses the selected provider.
Profiles can change speech and cleanup choices. Test a short dictation or meeting and inspect its status. Confirm all needed downloads are complete before going offline.
Local inference does not mean the app never connects to the network. Downloads, calendar access, optional services, and Trial and paid-license verification are separate. Read [Privacy and data flow](/providers/local-privacy) and [Release and access](/pricing/overview).
Personal notes and preparation questions guide insights and refinement and are included in those requests. Choose **On-device** to keep that processing local, including for Google-linked sessions. Choose local speech as well to process recorded audio on your Mac. Switching back from cloud processing or removing notes does not recall earlier transfers or sign out the provider.
# Troubleshooting
Source: https://docs.stenox.app/guides/troubleshooting
Recover permissions, model preparation, dictation, and Notetaker workflows.
## Stenox asks for a trial or license
Stenox requires an active trial or license to start new work. Open **Settings → License** to **Start 7-day trial**, **Activate** an existing key, or **Verify now** for paid access while online. Trial expiry is separate from the 30-day paid offline allowance. Viewing and exporting saved history remains available. See [Release and access](/pricing/overview).
If all device slots are in use, release an old Mac through **Manage purchase and devices**. Valid earlier paid licenses cover Stenox without another purchase and keep their original allowance. If an older key is rejected, contact support to check its status before buying again; inactive, refunded, revoked, or expired licenses do not qualify. A temporary network failure does not extend the cached verification deadline.
If the app says **Licensing is not configured**, it cannot start a trial or activate a key. Check that you have the intended build and contact support. Changing API keys or permissions will not configure licensing.
## Permissions or Keychain prompts after upgrading
Check that you opened Stenox from **Applications**. Review Microphone, Accessibility, and Screen & System Audio Recording under **System Settings → Privacy & Security**, and reopen the app if requested. Follow [Upgrading Stenox](/guides/upgrading) for the publisher name and older permission entries.
If Stenox reports that a saved credential needs access, use **Allow Keychain access** for the affected provider in **Settings → API Keys**, or the same button in **Settings → License** for licensing. Unlock your login Keychain if requested, and verify that the macOS request is for your installed Stenox copy before approving it.
Recovery keeps the saved credential and provider/model selections. Denying or cancelling leaves the credential unchanged; retry from Settings when ready. Licensing recovery does not start a trial or create a new activation. This flow is part of the [access recovery guide](/guides/upgrading#recover-provider-and-license-access).
Only add a provider key again if it is missing or you intend to replace it. Notetaker subscription sign-ins use their separate connection controls. If recovery fails, check for another running Stenox copy and send support the error wording with private details removed.
Do not erase Keychain, delete saved history, or allow all applications to access a credential. If a licensing identity was actually deleted or lost, activate your key again and release the old device slot through **Manage purchase and devices** if necessary. A denied access request alone does not require those steps.
## Recording does not start
Check that Stenox is running with an active trial or license, your input device is selected under **Microphone**, and Microphone permission is enabled in macOS. Wait for the selected speech provider to be ready. End any active meeting before dictating, or finish dictation before starting a meeting.
A language model is optional for dictation readiness. Do not troubleshoot a disabled Polish model as though it were the speech recognizer.
## A model is downloaded but not ready
Downloaded files and a loaded runtime are separate. Review the status in the relevant model controls and allow preparation to finish. Check internet access for missing files and free storage for the download. If preparation failed, use the offered retry after resolving the error.
Avoid deleting model folders by hand while Stenox is using them. Use [Settings → Models](/setup/models) to manage downloads and read any profile warnings.
## Text does not appear
Enable Accessibility for the installed Stenox app under **System Settings → Privacy & Security → Accessibility**. Put the cursor in an editable field and check whether your output mode copies to the clipboard. Some targets do not permit verified replacement; use Copy or retrieve the text from History.
If you installed a different signed build, check that permission is granted to the app you actually opened. Reopen it when macOS requests that step.
## Polish is unavailable or failed
Enable **Polish on demand** in Text cleanup and prepare the selected model or Gemini key. Polish is an explicit action after transcription. Missing local models are not downloaded implicitly by that action. A failure or cancellation keeps the dictated text.
If the field or clipboard changed, the result may be available to copy rather than safely replace. Fixed cleanup models ignore unsupported custom style controls; see [Local AI cleanup](/providers/llm/mlx-local).
## A meeting has missing or duplicated speech
Confirm both Microphone and Screen & System Audio Recording access. Current meetings capture both sources. Headphones reduce the chance that remote voices also enter the microphone. Check the active speech language and input signal.
Speaker separation is not echo cancellation. Overlap and background noise can affect words and attribution. [Edit or reassign passages](/meetings/conversation) rather than assuming the labels identify people automatically.
If microphone delivery stalls after an audio-device change, Stenox can show **Reconnecting audio…** and attempt to restart capture. Recovery is limited, and missing audio is not reconstructed. If the session is interrupted, check the device and start a new session; completed text is preserved.
## Insights or Ask stopped updating
Check the selected Notetaker provider, model, account, and reasoning choice. Account errors, limits, or timeouts can pause automatic reviews until explicit retry. Recording and saved text continue when optional insights fail.
A very short session may not yet have enough conversation: scheduled reviews need two contributions; **Generate insights** needs one and also works with **Automatic insights** off. Google-linked context uses the selected provider. After changing providers or models, pending old results are cancelled; wait for preparation and explicitly generate again if needed. Reference import supports readable text, not OCR of scanned pages. Check the coverage label for partial input.
## A saved session looks incomplete
Review the session's completion or interruption status. A recovered paused session cannot resume recording after the app restarts. Refinement failure does not erase accepted Current insights; export those if a refined draft is unavailable.
Personal notes, references, and Ask history are not bundled in the transcript exports. See [Export and keep your session](/meetings/exports).
If the problem persists, [send feedback](/support/feedback) with your app version, macOS version, selected provider/model, and the visible error. Remove keys and private content from any example.
# Upgrading Stenox
Source: https://docs.stenox.app/guides/upgrading
Upgrade Stenox and recover permissions or saved credentials.
## Choose when to upgrade
Upgrading from a free 1.x release is optional. The current app uses a **seven-day full-app trial**, followed by a license for new work. **Valid earlier paid Stenox licenses cover Stenox without another purchase**, with original rights and device allowances preserved. Inactive, refunded, revoked, or expired licenses do not qualify; see [Release and access](/pricing/overview).
## Before installing
Finish recordings and let saves complete. Export important history and make a current Mac backup before replacing the app. Keep your license receipt and access to your provider accounts available. Do not remove Stenox's saved data, models, or Keychain items to prepare for the upgrade.
## Update through Stenox
**The built-in updater is the normal upgrade path**. Follow Stenox's update prompt, or choose **Check for Updates...** in its menu when that option is shown. If the update is already downloaded, the menu can show **Restart to Update** instead. A separate manual download is not intended to be necessary for everyone.
If the updater cannot offer or complete the upgrade, use the manual fallback: download Stenox from the [Stenox website](https://stenox.app/download), quit Stenox, and install it in **Applications**. Open that installed copy.
After either path, check your history, profiles, shortcuts, input device, and model selections before relying on Stenox for a meeting. If something is missing, preserve your backup and [contact support](/support/overview).
## Check permissions again
macOS may ask again after the app's signing identity changes. In **System Settings → Privacy & Security**, check Stenox's access:
| Permission | Needed for |
| - | - |
| Microphone | Recording your voice for dictation and meetings. |
| Accessibility | Inserting dictated text into other apps. |
| Screen & System Audio Recording | Capturing the other side of a meeting. |
Grant access to the copy you opened from Applications and reopen it if macOS requests it. If Accessibility still refers to an older copy, remove that Stenox entry and add the installed app again. Follow Apple's [screen and system audio instructions](https://support.apple.com/guide/mac-help/control-access-screen-system-audio-recording-mchld6aa7d23/mac) for that permission. This is separate from activating a license.
## Recognize the publisher
Stenox's business name is **GKI Software Inc., formerly Sophosia Software Inc.** The expected Apple signing name for Stenox is **Sophosia Software Inc.**, under Team ID **VWXR46VR26**. macOS can therefore show Sophosia in a security or credential prompt even though the website uses GKI Software Inc.
Use the official download and check which app is requesting access. A matching display name alone is not a reason to approve an unexpected prompt.
## Recover provider and license access
An updated app can need permission to use an existing Keychain item. If Stenox reports that a provider's saved key needs access, open **Settings → API Keys** and choose **Allow Keychain access** for that provider. Confirm the macOS request is for your installed Stenox copy, and unlock your login Keychain if requested. Apple's [Keychain guidance](https://support.apple.com/en-asia/guide/keychain-access/kyca1331/mac) explains why an updated app may ask again.
This action updates access to the saved credential and reloads it without replacing or deleting its value. Your provider and model selections stay in place. If you deny or cancel the request, the saved credential stays unchanged and you can retry the same button later.
For licensing, use **Allow Keychain access** in **Settings → License** when offered. It restores access to the saved device identity and licensing state; it does not start another trial or create a new activation. Once access is restored, review the status shown there.
If no provider key is saved, add it under **Settings → API Keys**. Notetaker subscription accounts use their separate connection controls in **Settings → Notetaker**. If a licensing identity was actually deleted or lost, activation may be needed again; release the old device slot through **Manage purchase and devices** if necessary. Denied access alone is not a reason to delete a credential, start another trial, or release a device slot.
If recovery still fails, check whether another Stenox copy is running and [contact support](/support/overview). Do not erase Keychain or allow every application to read an item. Your paid offline allowance does not recover a missing local credential.
## Review changed models and Polish
Stenox inserts the transcript first. **Polish** is a separate action, and optional AI cleanup does not block speech-model readiness. Review **Text cleanup**, even if your older build rewrote dictation automatically.
**Gemma 3n 2B and 4B are no longer supported.** Saved selections for removed local models are normalized to the current default, **Gemma 4 E4B**. Check the selected model and its download before polishing. See [Local AI cleanup](/providers/llm/mlx-local) for the current and remaining legacy choices.
# Speak. Keep the words.
Source: https://docs.stenox.app/index
Dictate into your apps, capture meetings, and work with notes in Stenox.
Stenox brings dictation, meeting capture, and notes to your Mac. Choose local processing or connect a supported provider for the parts of your workflow you want to run in the cloud.
See [what’s new](/changelog), learn about [trial and license access](/pricing/overview), or follow [Upgrading Stenox](/guides/upgrading) if you already use Stenox.
Speak into a text field, then polish the result when you want to.
Follow the conversation, review decisions and tasks, and keep your own notes.
Find the speech, cleanup, and meeting models available in Stenox.
Set up profiles, shortcuts, and word replacements for your work.
## Dictation that stays in your flow
Place your cursor in an editable field, use your [dictation shortcut](/setup/hotkeys), and speak. Stenox transcribes your recording, applies your enabled filler removal and word replacements, then inserts or copies the text. **Polish** is a separate action after dictation; an AI model is not required to start recording.
## Meetings with a source to return to
[Notetaker](/meetings/overview) captures microphone and system audio, organizes a conversation by speaker, and optionally drafts takeaways, decisions, tasks, and open questions. Edit the transcript, follow citations, add reference documents, or [ask a question](/meetings/references-and-ask) about the session.
## Choose where processing happens
Local speech and language models run on your Apple Silicon Mac after download. Cloud speech sends audio; cloud cleanup and subscription insights send the text and context needed for the chosen feature. Your choices for speech, dictation cleanup, and meeting insights are separate. Read [Privacy and data flow](/providers/local-privacy) to set them deliberately.
# Install Stenox
Source: https://docs.stenox.app/installation
Install Stenox on Apple Silicon and prepare permissions and local speech.
## Check your Mac and build
The current app targets **Apple Silicon and macOS 15 or later**. Allow storage for each model you download, plus working memory for the models you run together. Download size is not a RAM requirement; Stenox does not publish a fixed per-model RAM guarantee.
If you already use Stenox, read [Upgrading Stenox](/guides/upgrading) before replacing it. See [Release and access](/pricing/overview) for trial and license details.
Visit the [Stenox download page](https://stenox.app/download). Email is optional. By entering your email and submitting the download form, you agree to occasional product updates and team check-ins. Close the prompt with **X** or press **Escape** to continue the download without an email. No account is created. See [download privacy](/providers/local-privacy#website-downloads) for details.
Open the downloaded DMG and drag Stenox to Applications.
Open Stenox from Applications and follow its introduction and permission setup. After an upgrade, review your saved settings and permissions.
Choose **Start 7-day trial** or activate your existing license while online. See [trial and license access](/pricing/overview). Existing free 1.1 installations do not require this step.
Enable Stenox under **System Settings → Privacy & Security → Microphone** and **Accessibility**. Microphone access records your voice; Accessibility supports inserting text into other apps.
New installations select **Parakeet v2 (English)**. Onboarding automatically starts preparing the selected local speech model, downloading missing files and loading cached ones. Keep Stenox open until it is ready; use **Try again** if preparation fails. Changing the local speech selection starts preparation for that model. Optional AI cleanup does not gate dictation readiness.
## Meeting permission
Notetaker also needs **Screen & System Audio Recording** access to capture the other side of a call. Current meeting setup captures microphone and system audio together. Grant the permission to the installed Stenox app and follow any macOS request to reopen it.
## First recording
Select your input in **Microphone**, open an editable text field, and try [your first dictation](/first-dictation). The default hold shortcut is **Fn/Globe**; **Option–Space** toggles recording. Saved shortcuts take precedence.
If setup fails, check model progress and free storage, then use the recovery steps in [Troubleshooting](/guides/troubleshooting).
# Calendar context
Source: https://docs.stenox.app/meetings/calendar
Connect Google Calendar and bring an event’s agenda and invitees into a session.
## Connect Google Calendar
Open **Settings → Notetaker** and connect Google Calendar through the browser sign-in flow. The integration reads events; it does not create or edit calendar entries.
The default calendar is your primary calendar. To add another calendar you can access, enter its **Calendar ID** from Google Calendar settings. Stenox does not request a full calendar-list permission.
The current Google grant requests only `calendar.events.readonly`. Stenox does not request Google profile access or read your account name and email to identify the connection. Organizer and invitee details, including names and email addresses, can still be read from events.
No connection is required to start a meeting. Manual sessions remain available.
## Choose an event
In **New session**, browse or search ongoing and upcoming events. The picker starts with a seven-day window and can load more. Select an event to bring its title, agenda, times, organizer, invitees, and available links into the session.
Requests cover the selected calendars and date window, with an explicit event-field selection. The saved event can contain more information than the smaller context sent for insights, including location, event and join links, attendee details, and attachment metadata.
The event description stays in its own context section and does not overwrite Personal notes. Starting consumes the draft event selection, so the next shortcut recording does not silently reuse it. A blocked start preserves the draft.
Invitee names can help you label speakers manually. When a name is absent, the participant picker shows the email address instead. Assigning that participant to a speaker can save the email as the speaker label, which can then appear in conversation inputs. These labels do not prove attendance or identify voices automatically.
## Choose how context is processed
Google-linked sessions and saved Google-derived context use your selected insights provider. Choose **On-device** for local insights. With **Codex**, **Claude Code**, or **Google Antigravity** selected, that provider can receive conversation, a Personal notes excerpt, included calendar and reference context, and information derived from them.
The speech provider is a separate choice. Select local speech too if you want meeting audio to remain on your Mac. Calendar context is background for the model, not evidence that someone said or agreed to an item.
Automatic reviews and **Generate insights** include the selected event's title (up to 500 characters), agenda (6,000), scheduled start/end and time zone, organizer name (120), and up to 100 invitee names (120 characters each). Dedicated email, location, event URL, join-link, and attachment-link fields are excluded from that calendar context. This does not remove emails used as speaker labels or personal information in free-form agenda, names, conversation, Personal notes, or reference text.
Refined notes and Ask use different inputs. They can include a calendar-derived session title or information already carried into outcomes, conversation, references, or provider context; they do not independently add the full event snapshot. See [the inputs for each operation](/meetings/processing#context-and-failures).
Attachment links can be opened in the browser. Stenox's Calendar permission does not include Drive file downloads. Download a file yourself and use [Add files](/meetings/references-and-ask) to import readable text.
## Reconnect and saved sessions
Saved event snapshots remain with the session after disconnecting. Reconnecting creates a new local connection identity and resets calendar selection. Older, broader grants may require reconnection; the app checks the returned scopes before use.
Disconnecting Calendar or removing its event does not erase information already carried into saved session content. Switching to On-device does not recall earlier cloud transfers or sign out the insights provider. Choose processing deliberately before generating; the app does not ask for fresh consent before each request.
**Outlook is not available in normal launches of this build.** Its implementation remains hidden pending publisher verification. A development acceptance switch is not public availability.
# Conversation and speakers
Source: https://docs.stenox.app/meetings/conversation
Review live conversation, edit words, and correct speaker attribution.
## Follow the live conversation
The **Conversation** tab groups consecutive passages from the same known speaker into a turn with a name and start time. Unknown speakers remain separate. Transcription arrives in windows, so the display may trail the live audio. Opening **Current insights** pauses Follow live so you can read without being pulled away.
Speech recognition and speaker attribution are estimates. Simultaneous speech, noise, missing timings, and quiet boundaries can produce missed words or incorrect assignments. Stenox does not separate overlapping voices into clean audio tracks.
## Name and reassign speakers
A speaker's pencil appears on hover or keyboard focus. Use it to rename that speaker or reassign the whole turn. Right-click an individual passage to reassign only that passage.
Renaming a speaker updates the displayed name and linked outcome owners while retaining the underlying source identity. If the session has calendar context, organizer and invitee names can be offered as choices. Choosing a name is a manual attribution; a calendar invitation is not evidence that the person attended or spoke.
Local diarization has up to eight arrival-order speaker channels. Speaker identities belong to the current meeting and are not voiceprints recognized across meetings. Cloud speaker labels are reconciled across overlapping requests and can remain uncertain at boundaries.
## Correct the transcript
Click a passage to edit its text. **Enter** or clicking away saves; **Escape** cancels. The original words and timing remain attached to the same source passage so citations still lead back to it.
A correction invalidates stale AI work. The next scheduled or manual insight review considers the corrected evidence. New speech continues in a separate passage while you edit, so a draft does not hide incoming words.
## Open source links
Insights and Ask answers can link to original conversation passages. Follow the link and review the surrounding turn before relying on a task, decision, or answer. A valid citation establishes a source location, not a guarantee that the AI interpretation is correct.
Transcript exports use your current edits and speaker names. An already exported file remains a snapshot; export again to include later corrections. [Export formats](/meetings/exports) explains timing and content boundaries.
# Export and keep your session
Source: https://docs.stenox.app/meetings/exports
Download transcripts and Markdown notes, with clear limits on what each file contains.
Use the session's download menu to choose the content you need. Transcript and note exports are separate.
| Download | Contains |
| - | - |
| **Transcript (.srt)** | Timed subtitle cues and speaker names. |
| **Transcript with timestamps (.txt)** | Consecutive same-speaker turns with start/end times. |
| **Refined session notes (.md)** | The saved refined Markdown draft, including your edits. |
| **Current structured notes (.md)** | Accepted takeaways, decisions, tasks grouped by owner, and questions, with statuses and deadlines. |
## Review before sharing
Transcript formats use the current edited words and speaker names. SRT keeps the recording timeline; an edited passage uses its original interval because new word alignment is not known. Transcript files do not include insights.
Current notes can be rendered directly from accepted outcomes. Refined notes need a successfully prepared or manually edited draft. If refinement fails, use Current structured notes rather than assuming an empty or old draft contains the whole session.
Downloading saved notes, reopening them, or switching Raw / Formatted display does not call the AI provider. Requesting a new refinement does.
## What is not in these exports
These download options are not a full session backup. They do not bundle the Personal notes field, the entire Ask history, imported reference files, or a complete editable session database. Insights and refined notes can reflect the priorities in your Personal notes, but exporting them does not export the original personal writing. Copy that writing separately when you need it outside Stenox.
The app stores session data locally, including saved extracted reference text. Its active SQLite storage is plaintext; do not describe local storage as app-encrypted. Use your Mac's own access and disk-protection settings as appropriate.
## Revisit and edit
Search Notetaker history and open the session to keep working. Later edits do not update files already downloaded; export a new snapshot when needed. Session deletion requires confirmation and is unavailable during active capture. Save or export the material you want to keep before deleting a session.
# Insights and personal notes
Source: https://docs.stenox.app/meetings/notes
Review takeaways, decisions, tasks, questions, and refined notes without losing edits.
## Current insights
With **Automatic insights** enabled, Stenox reviews new or corrected conversation on the meeting clock at roughly two-minute intervals and performs a final review when the session ends. Scheduled reviews need at least two conversation contributions. Processing can take longer than the interval.
Choose **Generate insights** to request a review now, including during capture and when automatic insights are off. It needs at least one conversation contribution and an available model. Turning automatic insights off preserves existing insights and edits; it does not change the selected provider or disable explicit generation.
Open the **Insights model** menu beside Generate insights to change the provider, model, reasoning choice, or automatic setting. Pending work from the old selection is cancelled and stale results are rejected while recording continues. See [providers and data flow](/meetings/processing) before selecting cloud processing.
Current insights groups **Takeaways**, **Decisions**, **Tasks**, and **Open questions**. In Conversation, it starts collapsed at the bottom with a count. The **Insights** tab also offers **Current** and **Refined** views.
Add or edit an outcome, change its type, set an owner or deadline, mark completion, remove it, or inspect its sources. User edits and removals are protected from later AI updates. Unknown owners remain unassigned rather than becoming invented commitments.
Insights are drafts. Review their evidence, names, dates, and ownership. The model works with bounded recent and changed conversation plus existing outcomes; it does not re-read an unlimited transcript on every update.
## Open-question prompts
During recording, eligible generated questions can appear one at a time in the notch or top floating surface. Hover pauses the reading timer; **Dismiss question** clears the prompt, and **Next question** appears when another is waiting. VoiceOver uses explicit dismissal.
Dismissing a prompt removes it from the session's display queue, not from the notes. Acknowledged questions do not repeat. Pausing, disabling insights, or ending the meeting suppresses the prompts; reviewing saved history does not replay them.
## Refined notes
After the final review, the selected insights provider can consolidate accepted outcomes into a separate Markdown draft. Refined notes do not replace the cited Current insights. A session with no user or calendar title may receive a suggested title from this pass; existing titles are preserved.
This pass uses the session title, accepted visible outcomes with owners, dates and status, included references, and an excerpt of Personal notes. Your notes guide the order and emphasis of accepted outcomes; they are not an independent source of facts. The request does not independently add the full transcript or calendar event. Its inputs and continued provider context can still contain calendar-derived information.
Refined text is editable and autosaves. If you edit it, a later regeneration produces a suggestion that you explicitly accept or discard. It does not overwrite your writing. Evidence changes can mark an unedited draft stale; review or regenerate before sharing it.
Use **Raw / Formatted** to switch the display of Current or Refined notes. Current's Raw view shows its export Markdown; edit individual insights in Formatted. Refined's Raw view is the Markdown editor. Switching display does not call AI.
## Personal notes
Personal notes autosave independently, including before recording starts. Use them for preparation questions and priorities you want insights to focus on. The editor shows **Saved on this Mac · Used to guide insights**.
**Personal notes are included in insight and refinement requests to your selected provider.** On-device processes them locally; a selected cloud provider receives the included text. Long notes use an excerpt based on a 6,000-byte UTF-8 limit, followed by an omission marker. The full note stays saved on your Mac.
Preparation guides relevance, not evidence: answers and task assignments still need support from the conversation. A proposal, assumption, or pasted AI answer in your notes does not establish an agreement. Refined notes emphasize relevant accepted outcomes without treating your preparation as new facts.
Adding, changing, or removing notes makes prior reviews eligible for reconsideration and invalidates an unedited refined draft. Results based on older notes are rejected, and continued provider context is reset for changed input. Your accepted edits remain protected. Use **Generate insights** to request another review when ready. Removing notes cannot retract text already sent to a cloud provider.
The editor supports Markdown headings, emphasis, lists, tasks, quotes, links, and code. Syntax appears in the paragraph you are editing and hides again when you leave it. Return continues a list or unchecked task; Return on an empty item leaves the list. Command–B and Command–I format a selection, with native undo.
## Change the session prompt
Open **Settings → Notetaker → Advanced** for a custom session prompt of up to 4,000 characters. Specify focus, language, tone, or detail. Save it to apply to future reviews, or choose **Use Stenox default**. Evidence and citation validation still apply.
If AI is unavailable, recording and saved text continue. You can still keep personal notes and use accepted Current insights. See [processing and recovery](/meetings/processing) and [exports](/meetings/exports).
# Meetings with Notetaker
Source: https://docs.stenox.app/meetings/overview
Start a session, capture microphone and system audio, and return to saved meetings.
Notetaker keeps the conversation, generated insights, and your personal notes together. It is included in [Stenox](/pricing/overview).
## Start a session
Open **Notetaker** in the main window. When idle, it shows session history. Choose **New session** for quick setup.
Check the speech model and insights provider. Optionally select a calendar event, add reference files, or write Personal notes and preparation questions to guide insights. An excerpt of those notes goes to the selected insights provider. Speech follows the active dictation model; insights have their own selection.
Choose **Start session**. Grant Microphone and Screen & System Audio Recording access if prompted. Stenox prepares speech and speaker processing, opens the session tracker, and minimizes the main window.
Use **Conversation**, **Insights**, and **Personal notes** as the meeting progresses. Set or edit the title in the tracker. Wear headphones on calls to reduce duplicate microphone and system audio.
The default **Control–Option–M** shortcut starts or ends a session from another app. It can start directly without opening quick setup. Change it in **Settings → General**. Dictation and meeting capture are mutually exclusive.
Current session setup records **microphone + system audio** together. Stenox does not join a call as a bot or read participant names from the call app. Make sure participants understand that you are recording.
## Pause, return, and finish
**Pause** stops both audio sources and commits queued speech. It retains the session, speaker names, and processing state. **Resume** continues the same recording clock without counting the pause. Preparation and finalization cannot be paused.
Closing the tracker window leaves capture running. Reopen Notetaker to return to it. The window can sit alongside another app; clicking that app does not end the session.
End the session with its stop control or the meeting shortcut. Stenox finishes pending transcription, performs a final insight review when enabled, and prepares refined notes when possible. A preparation or AI failure is shown in the session; review the status before assuming notes are complete.
## Find an earlier session
History searches titles, conversation, insights, personal notes, and extracted reference text. Double-click a row or press Return to open it. During active capture, an older session opens as a read-only snapshot; when idle, saved sessions are editable.
A paused session recovered after the app quits is saved as interrupted history. Recording cannot resume across app launches. Quitting normally waits for capture finalization and pending saves.
Correct words and attribution without losing source links.
Review drafts, protect edits, and choose what to export.
Bring context and inspect cited answers.
Configure speech, local insights, and subscriptions.
# Meeting models and processing
Source: https://docs.stenox.app/meetings/processing
Configure shared speech, local speaker separation, and independent insight providers.
## Three jobs, separate choices
| Job | Current behavior |
| - | - |
| Speech to text | Uses the active dictation provider and model, including profile selection. |
| Speaker separation | Local Nemotron where needed, or timed speaker labels from supported cloud speech adapters. |
| Insights, refined notes, Ask | On-device Qwen 3.5 4B, or a selected subscription integration. |
Quick setup's speech menu changes the shared speech model; there is no separate meeting-only speech override. A ready local speech runtime is reused. While a meeting holds it, competing jobs and model changes are guarded. Pause retains it; ending releases the meeting reservation without unloading dictation's selected speech model.
## Local meeting models
**On-device** insights use **Qwen 3.5 4B** (`qwen3_5_4b_4bit`, catalog estimate \~3.1 GB). This is independent of the model selected for dictation Polish and uses a separate inference instance.
For local speech and providers without speaker labels, Stenox loads **Nemotron-3-Diarization-8bit** for meeting-local speaker separation. It is not a dictation cleanup model and does not appear as an interchangeable speech model. Its state is scoped to one meeting and is released when that meeting finishes.
Initial preparation may download these files. No fixed RAM minimum or processing speed is guaranteed. Local speech, diarization, and insights together need more resources than speech alone. Turning automatic insights off reduces the work while retaining transcription and personal notes.
## Cloud speech behavior
Deepgram, AssemblyAI, and Gemini use timed meeting transcription with speaker requests. Other adapters use their normal file transcription path, with local speaker handling where needed.
* **Deepgram** follows the selected Nova model and language. Streaming-only Flux uses Nova 3 for meeting audio windows, as it does for file transcription. Requests opt out of model improvement.
* **AssemblyAI** uses Universal-3.5 Pro with its Universal-2 fallback and selected region/language. Stenox requests deletion of completed jobs.
* **Gemini Transcribe** uses Gemini 3.5 Transcribe with Interactions storage disabled.
These request settings do not promise zero provider retention. Cloud speech sends audio to that provider. Local windows are approximately ten seconds; cloud windows are approximately thirty seconds, with overlap. Transcription is not instantaneous and speaker identity may be uncertain across requests.
## Subscription insights
In **Settings → Notetaker**, choose **Codex**, **Claude Code**, or **Google Antigravity**. Use **Sign in** if needed, then choose an available model and reasoning level. Models come from the connected runtime; Stenox does not promise a fixed list or silently replace a missing saved model.
Codex and Claude Code use bundled runtimes; they do not require a separate Node, Bun, or CLI installation. Existing eligible coding-runtime sign-ins can be reused. These integrations require subscription authentication, not an API key. Provider limits and usage apply.
Google Antigravity requires its official `agy` CLI installed separately. This adapter accepts 1.2 releases starting at 1.2.12. Its UI identifies the active CLI login source, not a verified email address.
Selecting another insights provider or turning insights off does not sign you out of the external provider. Stenox does not expose a separate Sign out or Disconnect control for these subscription accounts.
The session's **Insights model** menu also lets you change the provider, model, reasoning choice, and **Automatic insights** setting during capture or while editing saved history. Switching cancels pending insight work and rejects stale results while capture continues. Existing insights and your edits remain. Returning to On-device affects future work; it cannot recall earlier transfers.
## Context and failures
Your selected provider handles Google-linked and manual sessions. With a cloud provider selected, the inputs below leave your Mac; **On-device** processes them locally.
| Operation | Newly assembled input |
| - | - |
| Automatic reviews and Generate insights | Session title, bounded changed/recent conversation, speaker IDs and names, current outcomes, elapsed duration, new contribution IDs, included reference text, bounded selected-event context, and a Personal notes excerpt. |
| Refined notes | Session title, whether to suggest a title, accepted visible outcomes with owner names, dates and status, included reference text, and a Personal notes excerpt. It does not independently add the transcript or full event snapshot. |
| Ask Stenox | Your question, session title, bounded selected conversation with speaker labels, included references, and up to three prior question/answer pairs. It does not independently add Personal notes or the full event snapshot. |
Personal notes guide focus in insights and refinement. Long notes use an excerpt based on a 6,000-byte UTF-8 limit plus an omission marker; the full saved note is unchanged. Notes do not establish facts, commitments, or owners. Editing or removing them invalidates stale results and refreshes the relevant provider context.
Titles and saved content can contain information derived from [Google Calendar](/meetings/calendar). Codex and Claude Code can continue provider context across Insights, refinement, and Ask for the same session and selection; that context can contain Personal notes from earlier turns. Provider caching is not guaranteed. Antigravity sends the selected context afresh each time. A request's listed inputs do not imply that a continued provider session has forgotten previous inputs.
Provider account terms, usage, retention, and training settings still apply. Selecting an integration does not establish a zero-retention or no-training guarantee.
Invalid output, account errors, timeouts, or usage limits preserve recording and saved text. They may pause automatic reviews until an explicit retry. **Generate insights** remains an explicit action when automatic insights are off. There is no silent switch to another provider or paid API account.
An audio capture or processing overrun can end a session; completed text is preserved. Check the status, export what is available, and resolve the input or resource problem before starting again.
# References and Ask Stenox
Source: https://docs.stenox.app/meetings/references-and-ask
Add readable documents, ask session questions, and inspect the cited evidence.
## Add reference documents
Open **References** in session setup or the tracker. Use **Add files** or drag files into the reference view. Supported files are **PDF, DOCX, PPTX, UTF-8 TXT, and Markdown**.
Each file is limited to 20 MB, 500 pages or slides, and 2 MB of extracted text. A batch accepts up to 20 files. Duplicate content is detected by its hash. You can preview, exclude, or remove an imported reference; removing it does not delete the original file.
Only readable text is extracted. Scanned pages, images, charts, and document layout are not interpreted. DOCX body text is stored as one section; PPTX text is split by slide. A scanned PDF needs a readable text layer before its content can help the session.
## What the AI receives
Included references are background context, distinct from evidence of what people said. Stenox uses full text when it fits the selected model's budget and otherwise chooses relevant page excerpts. The UI indicates partial coverage. A large model context window does not guarantee every document fits.
Excluded references are not included in future requests. Changes to included evidence invalidate stale replies and reset continued provider context where needed. They cannot retract text already sent to a cloud provider.
[Google Calendar context](/meetings/calendar) follows your selected insights provider, including a cloud provider. Calendar attachment links do not grant Stenox permission to download the files automatically; open them in your browser and add a downloaded file explicitly.
## Ask a question
Open **Ask Stenox** from the session. It appears beside a wide workspace and in a sheet in a narrow one. Ask about the conversation or included references, then follow the answer's source links to the passage or stored reference page.
Ask uses the same provider, model, account, and reasoning choice as Insights. Its input includes your question, session title, bounded selected conversation with speaker labels, included reference text, and up to three prior question/answer pairs. It does not independently add the full calendar event snapshot. The title, other inputs, or continued provider context can still contain Google-derived information. Read the coverage shown with the answer; it may not cover every part of a long meeting.
## Decide what to keep
An answer does not automatically change your notes. Choose **Save to personal notes** or **Add to insights** when you want to keep it. Otherwise it remains a question-and-answer result.
Ask does not independently add the Personal notes field to its request. A continued Codex or Claude Code session can contain notes from an earlier insight or refinement turn, so do not treat that omission as a privacy boundary or assume Ask has read your latest notes. Ask answers still require conversation or reference evidence.
Choosing **Save to personal notes** makes that text eligible for later insight and refinement requests. Saved AI text remains background, not evidence of an agreement. Review answers against their sources before turning them into commitments or sharing them.
# Release and access
Source: https://docs.stenox.app/pricing/overview
Current download availability, Stenox's full-app trial, and paid license access.
## App access
Stenox includes a seven-day full-app trial, followed by a license for new work. See [what’s new](/changelog) or visit the [Stenox website](https://stenox.app) for downloads and current purchase options.
Existing free 1.1 installations remain usable without a license. Upgrading to the current app is optional. Older screens, models, and features can differ from these guides; version 1.1 does not include Notetaker or licensing controls. See [Upgrading Stenox](/guides/upgrading).
## License access
New licenses cover **two Macs** and permanent access to purchased features. There is no app subscription; future major upgrades may cost extra. Optional third-party provider usage is separate.
## Start the seven-day trial
In Stenox setup or **Settings → License**, choose **Start 7-day trial** while online. No account or card is required. The trial starts when that request succeeds, not when you download the app, and gives access to the whole app for seven days.
After the trial ends, activate a license to start new work. There is no ongoing free tier. You can still view and export saved history, and work already in progress can finish. The paid offline allowance does not extend a trial.
## Activate a new or existing license
Open **Settings → License**, enter your Stenox license key, and choose **Activate** while online. Your key and receipts are available through **Manage purchase and devices**.
**Valid earlier paid Stenox licenses cover Stenox without another purchase**, with their original rights and device allowance preserved. An existing allowance is not reduced to the new two-Mac limit. Inactive, refunded, revoked, or expired licenses do not qualify. If an older key is rejected, [contact support](/support/overview) to check its status before buying again. An earlier free download is not a purchased license.
## Verify before going offline
Paid licenses refresh every **seven days** and allow up to **30 days offline from the last successful verification**. Use **Verify now** in **Settings → License** before an extended offline period. Local models must also be downloaded and ready; cloud providers still need internet access.
A temporary connection failure preserves already verified access until its existing deadline; it does not restart the 30 days. After that deadline, reconnect and verify to start new work. History, exports, and work already in progress remain available. A confirmed revoked license or removed activation can stop new work sooner. Any expiry in the original purchase terms still applies.
## Change Macs
In **Settings → License**, choose **Deactivate this Mac** and confirm while online to release its activation. Saved work stays on the Mac. If you no longer have that Mac, use **Manage purchase and devices** to release its slot, then activate the replacement.
If a saved licensing credential needs permission, first use **Allow Keychain access** in **Settings → License**. This restores access without starting a trial or creating a new activation. If the licensing identity was actually deleted or lost, the same Mac may need activation again; release the old activation if it occupies a slot. Use [Troubleshooting](/guides/troubleshooting) if you cannot identify it. Do not reset Keychain as an activation fix.
## Local processing and account access
Local model inference does not require an API key after model download. Cloud transcription and Gemini cleanup use your own provider keys. Notetaker subscription insights use an eligible connected account, with that provider's usage limits. App access and provider billing are separate.
# Automatic profile selection
Source: https://docs.stenox.app/profiles/auto-activation
Use app and browser-title rules without assuming priority or changing a manual choice.
## Turn on automatic mode
Select automatic profile mode when you want Stenox to match the active app. Selecting a profile manually keeps that choice in effect until you return to automatic mode.
## Rule types
| Rule | Matching behavior |
| - | - |
| App Bundle ID | Exact match to the active app's identifier. Use the app picker to avoid guessing it. |
| Window Title Keyword | Case-insensitive substring match against the detected browser window title. |
A title keyword is not a URL rule, regular expression, or proof of which website is open. It depends on the title available to Stenox and Accessibility access. Avoid broad keywords that match unrelated tabs.
Profiles are checked in their existing order and the first eligible match wins. Any matching rule can select its profile; app rules do not have a separate guaranteed priority over title rules. If nothing eligible matches, Stenox falls back to Default.
## Verify your rules
1. Prepare the profile's selected models or provider keys.
2. Enable automatic mode.
3. Switch to the target app or browser tab and check the active profile.
4. Switch to an unrelated app and check that the expected fallback is selected.
5. Record a short sample before relying on the setup.
If an unexpected profile wins, narrow overlapping rules. If no profile activates, check that automatic mode is on, the model is available, the app identifier matches, and the browser title contains your keyword.
# Create a profile
Source: https://docs.stenox.app/profiles/creating-profiles
Save and verify a profile with available speech and cleanup models.
Create a new profile and give it a name that describes its purpose, such as English notes or French messages.
Select a current provider, model, and supported language. Download local models or configure the provider's API key before relying on the profile.
Set filler removal independently. Enable optional Polish and select its provider/model if you want an AI action after dictation. Add custom instructions only for models that support them.
Choose **Save Profile**, select it, and dictate a short sample. Check the active model and the result before using it for longer work.
## Prepare dependencies
Use **Settings → Models** for downloads and **Settings → API Keys** for supported cloud services. An unavailable provider or missing model is a setup problem, not permission for Stenox to silently send your data somewhere else.
The [provider overview](/providers/overview) lists current choices. Deprecated models may remain visible for migrated selections; use the current catalog for new profiles.
## Add automatic rules
Add an app rule using the app picker, or a window-title keyword for a browser context. Enable automatic profile selection to use the rules. [Automatic selection](/profiles/auto-activation) explains matching and fallback.
Default follows the app's global settings. Edit those settings when you want to change the baseline; use custom profiles for distinct configurations.
# Profiles
Source: https://docs.stenox.app/profiles/overview
Save speech and cleanup settings for different apps and kinds of work.
A profile combines a speech provider, model, language, filler-removal setting, and optional Polish configuration. Use separate profiles when your language or writing needs change between apps.
## Default and custom profiles
**Default** reflects the app's current settings. It is not a separate stored preset that you can delete. Custom profiles save their own choices, including provider/model, language, whether Polish is enabled, and supported instructions.
Open **Profiles** to create or edit a profile. Pick a profile manually when you want it to stay selected. Choose automatic mode when you want matching trigger rules to select it for you.
## What a profile changes
| Setting | Effect |
| - | - |
| Speech provider, model, language | How audio becomes text. |
| Remove filler words | Applies deterministic filler removal independently of AI. |
| Polish settings | Chooses the provider/model and applicable instructions for an explicit Polish action. |
| Trigger rules | Match an app bundle ID or a browser window-title keyword in automatic mode. |
Profiles do not make AI cleanup automatic. Stenox delivers dictation first, then you choose **Polish**. Fixed-purpose models such as S1-mini have their own instruction limits.
Notetaker follows the active speech selection, including a profile. Its insight provider remains separate. Changing a speech model is guarded while a meeting is using it.
Choose providers and prepare the models you need.
Match apps and browser titles deliberately.
# Writing instructions
Source: https://docs.stenox.app/profiles/system-prompts
Configure optional Polish instructions and distinguish them from meeting prompts.
Writing instructions guide **Polish**, the optional AI action after dictation. They do not change the audio recognizer or automatically rewrite every recording.
## Add a focused instruction
For a general cleanup model, open the profile editor or **Text cleanup** and set the supported writing-style or custom-instruction controls. Keep the instruction specific and check the result on representative dictation.
For example:
```text theme={null}
Use short paragraphs. Preserve names, numbers, links, and the speaker's intent.
```
An instruction is a preference, not a guarantee of factual preservation. Review important text before sending it.
## Model limits
**S1-mini by Superwhisper** uses a fixed English cleanup pass and semi-formal style. Writing style, app overrides, and custom prompts do not apply. **Yooz-Light v3** also uses a specialized English correction path. Select a general model for broader writing instructions; see [Local AI cleanup](/providers/llm/mlx-local).
A prompt cannot add a language or capability that the selected model does not support. Word or phrase substitutions belong in [Vocabulary](/setup/vocabulary).
## Meeting prompts are separate
Use **Settings → Notetaker → Advanced** to set the focus, language, tone, and detail of future session insights. That prompt has a 4,000-character limit and retains source validation and protection for user edits. It is independent of the dictation profile's writing instructions.
# Cloud providers
Source: https://docs.stenox.app/providers/cloud-byok
Connect supported cloud speech and cleanup services with your own keys.
Bring your own key when you want a supported cloud service. Provider availability in Stenox and your account's entitlement are separate; a configured key does not include free usage or guarantee access to every model.
| Provider | Speech | Dictation Polish |
| - | - | - |
| Deepgram | Yes | No |
| AssemblyAI | Yes | No |
| OpenAI | Yes, OpenAI Transcribe | No |
| Google Gemini | Yes, Gemini Transcribe | Yes |
The current catalog does not offer Groq or Stenox-hosted cloud processing. Choose a listed provider if an older profile references a removed service.
## Configure deliberately
1. Create a key in the provider's own console.
2. Add it under **Settings → API Keys** in Stenox.
3. Choose the provider in your speech profile or Text cleanup settings.
4. Try a short sample and check the provider account for access, limits, and usage.
Speech sends audio. Polish sends text and applicable instructions. A local speech provider with cloud Polish still sends text when you click Polish. See [Privacy and data flow](/providers/local-privacy).
## Recover access to a saved key
If a provider reports a Keychain access error, choose **Allow Keychain access** on its card in **Settings → API Keys**. Confirm the macOS request is for your installed Stenox copy and unlock your login Keychain if requested.
Recovery authorizes the saved item for Stenox and reloads its existing key without replacing or deleting it. Your provider and model choices are preserved. If you deny or cancel, the key stays saved and you can retry from the same card. Add a key again only if none is saved or you intend to replace it. This flow is part of the [access recovery guide](/guides/upgrading#recover-provider-and-license-access); that guide also covers licensing and other account recovery.
## Provider accounts and usage
Provider pricing, credits, quotas, regions, and retention can change. Check the provider's current console and documentation; the old fixed credit and per-minute figures are not maintained here. Stenox app access does not pay for those services.
For meeting insights using Codex, Claude Code, or Antigravity, use [Notetaker subscription setup](/meetings/processing). Those are separate from BYOK dictation providers.
# Gemini cleanup
Source: https://docs.stenox.app/providers/llm/google-gemini
Use an API key for optional on-demand text polishing.
Gemini is the current cloud provider for dictation **Polish**. Add a Gemini API key under **Settings → API Keys**, enable **Polish on demand** in **Text cleanup**, and choose **Gemini**.
## Models configured in this build
* **Gemini 3.5 Flash-Lite** — `gemini-3.5-flash-lite`, the new default Gemini selection.
* **Gemini 3.5 Flash** — `gemini-3.5-flash`.
* **Gemini 3.7 Flash** — `gemini-3.7-flash`.
These labels reflect Stenox's current catalog. Account entitlement and successful inference still depend on the provider. The app does not silently upgrade an unavailable model or include provider usage in its own price.
## Use Polish
Dictate normally and wait for the transcript to be inserted or copied. Click **Polish** to send the text and applicable writing instructions to Gemini. The original remains in History. If the request fails, the dictated text is retained.
A local speech model with Gemini Polish is a mixed setup: speech stays local, but the Polish action sends text to Google. See [Privacy and data flow](/providers/local-privacy).
Gemini speech uses a separate [Gemini Transcribe](/providers/transcription/gemini-transcribe) selection. Meeting insights use [Notetaker's providers](/meetings/processing), not this dictation cleanup menu. Review current quotas, access, and policies in [Google AI Studio](https://aistudio.google.com).
# Local AI cleanup
Source: https://docs.stenox.app/providers/llm/mlx-local
Use on-demand local Polish and understand current and legacy cleanup models.
Local cleanup runs on Apple Silicon using MLX. In **Text cleanup**, enable **Polish on demand**, choose **Local**, and prepare a model. Click **Polish** after dictation or file transcription to use it. Selecting a model does not automatically rewrite every recording.
## Current cleanup catalog
These are the models offered for a new selection in the reviewed app build. Sizes are catalog download estimates; they are not RAM requirements or benchmark results.
| Model | ID | Estimated download | Behavior |
| - | - | - | - |
| S1-mini by Superwhisper | `s1_mini` | \~633 MB | Fixed English cleanup; includes punctuation, capitalization, and spoken self-corrections. |
| Yooz-Light v3 | `yooz_light_v3` | \~612 MB | English dictation correction with bounded, validated output. |
| Qwen3.5 4B | `qwen3_5_4b_4bit` | \~3.1 GB | General cleanup with the selected writing instructions. |
| Gemma 4 E2B | `gemma4_E2B_it_4bit` | \~3.6 GB | General cleanup with the selected writing instructions. |
| Gemma 4 E4B | `gemma4_E4B_it_4bit` | \~5.2 GB | Default local cleanup selection; Polish starts disabled. |
S1-mini uses a fixed semi-formal style. Writing-style controls, app overrides, and custom instructions do not apply to it. It has a bounded input and generation budget; unsupported or invalid output keeps the original text. Yooz-Light also uses its specialized cleanup path. Use a general model when your workflow needs custom writing instructions.
Model availability in the app catalog is not a guarantee of a particular speed or rewrite quality. Review important details. No model is documented here as learning your vocabulary automatically.
## Load before polishing
Download and load the selected model through **Text cleanup** or pre-download it in **Settings → Models**. **Polish** can wait for an in-progress warm-up, but it does not silently download missing files or fall back to a cloud service. Speech recording remains independent of optional cleanup readiness.
The original dictation remains in History. If polishing fails or is cancelled, the dictated text stays available. If the original insertion can no longer be verified, copy the proposed result rather than expecting Stenox to replace unrelated text.
## Older models
These deprecated models can still load when already selected. They are not part of the normal new-selection list.
| Model | ID |
| - | - |
| Qwen 2.5 1.5B | `qwen2_5_1_5B_4bit` |
| Phi-3 Mini | `phi3_mini_4bit` |
| Qwen 3 4B | `qwen3_4b_4bit` |
**Gemma 3n 2B and 4B have been removed.** Saved local selections for unsupported models are normalized to **Gemma 4 E4B**, the current default. Review the selection and prepare that model before polishing. **Gemma 4 12B** is not in this build's supported cleanup catalog.
An older profile is not automatically a recommendation. Move to a current model deliberately and compare the result on your own text.
## Meetings are separate
The On-device option in Notetaker uses **Qwen 3.5 4B** for insights, refined notes, and Ask Stenox. That setting is independent of the dictation cleanup selection. [Meeting processing](/meetings/processing) explains its lifecycle and additional speaker model.
# Privacy and data flow
Source: https://docs.stenox.app/providers/local-privacy
Understand what stays on your Mac and what each optional provider receives.
Local-first describes a processing choice, not a promise that every feature is offline. Configure speech, Polish, and Notetaker insights separately.
| Feature choice | Processing and data flow |
| - | - |
| Parakeet, Whisper, Cohere Transcribe | Audio is processed on your Mac after model download. |
| Local Polish | Dictated text is processed by the selected local MLX model. |
| On-device meeting insights | Conversation, a Personal notes excerpt, and included calendar/reference context are processed locally with Qwen 3.5 4B. |
| Cloud speech | Audio is sent to the selected speech provider. |
| Gemini Polish | Text and the configured writing instructions are sent to Gemini. |
| Subscription insights, refinement / Ask | The selected Codex, Claude Code, or Google Antigravity integration receives the [inputs for that operation](/meetings/processing#context-and-failures), which can include Google-derived information. |
| Google Calendar | Read-only event access uses the connected Google account. Included context follows the selected local or cloud insights provider. |
| Trial and paid activation | Online requests verify app access and the device; they do not upload audio, transcripts, or notes. |
## Set up local processing
Choose a local speech provider, choose **Local** for any enabled Polish, and choose **On-device** for Notetaker insights. Download required models while online, then verify they are ready before relying on offline inference. A profile can override speech or cleanup choices, so inspect the active profile too.
Personal notes and preparation questions are included in insight and refinement requests to the selected provider. Long notes use a bounded excerpt; the full saved note stays on your Mac. Ask does not add that field directly, but continued provider context may retain it from earlier turns. See the [operation inputs](/meetings/processing#context-and-failures).
Removing Personal notes, excluding a reference, disconnecting Calendar, removing an event, or switching back to On-device cannot retract earlier cloud submissions. Saved titles, outcomes, and other session content may retain derived information.
Turning **Automatic insights** off stops scheduled reviews; **Generate insights** still uses the selected provider when you choose it. Switching providers does not sign out their accounts. See [Calendar context](/meetings/calendar) for Google's events-only permission, fetched fields, and bounded insight context.
## Website downloads
Email is optional. By entering your email and submitting the download form, you agree to occasional product updates and team check-ins. You can ask the team to stop these emails at any time. Close the prompt with **X** to continue the download without an email. No account is created.
The website records download starts with or without an email; these are not counts of completed downloads or installations. If you submit an email, it is associated with that download record and kept out of analytics events. This prompt appears only on website downloads; in-app updates continue through the updater.
## Storage and services
History and sessions are stored locally. **Active SQLite databases are plaintext**, not encrypted by Stenox. Local storage is distinct from your Mac's access controls and disk encryption.
Provider API keys use macOS Keychain. Subscription runtimes own their sign-in credentials. Provider retention, usage, and account policies still apply to cloud processing; opt-out or deletion request flags are not a zero-retention guarantee.
Licensing credentials also use Keychain. Starting the trial and activating a license require internet access. Paid access refreshes weekly and can remain offline for up to 30 days from the last successful verification. The trial ends after seven days, without the paid offline allowance.
Downloads, optional integrations, and app services have their own network requirements. Offline local inference is separate from [app access](/pricing/overview). Read the [Stenox privacy policy](https://stenox.app/privacy) for the public policy.
# Choose your providers
Source: https://docs.stenox.app/providers/overview
Separate speech recognition, optional text cleanup, and meeting insights.
Stenox has three independent processing choices. A local speech model does not make cloud cleanup or subscription insights local.
| Stage | Local choices | Cloud choices |
| - | - | - |
| Speech recognition | Parakeet, Whisper (WhisperKit), Cohere Transcribe | Deepgram, AssemblyAI, OpenAI Transcribe, Gemini Transcribe |
| Dictation Polish | Local MLX cleanup models | Gemini with your API key |
| Meeting insights, refined notes, Ask Stenox | On-device Qwen 3.5 4B | Codex, Claude Code, Google Antigravity subscription integrations |
## Speech recognition
New installations use **Parakeet v2 (English)**. Speech is selected through the Default profile or a custom profile. Notetaker follows the active speech provider and model, including a profile selection; its quick setup model menu changes that shared selection.
English and multilingual local speech.
A range of local model sizes through WhisperKit.
Local MLX speech with an explicit language selection.
Supported audio services using your own keys.
## Optional text cleanup
Stenox inserts the transcript before AI rewriting. Turn on **Polish on demand** in **Text cleanup**, select and prepare a model, then click **Polish** after dictation or file transcription. See the complete [local cleanup catalog](/providers/llm/mlx-local) and [Gemini setup](/providers/llm/google-gemini).
## Meeting models
Notetaker reuses speech recognition but manages speaker separation and insights separately. [Meeting processing](/meetings/processing) explains the additional downloads, subscription setup, and data sent for each operation.
## Model downloads and older settings
[Manage local models](/setup/models) explains downloaded versus loaded state. Retired choices may remain in migrated profiles; they are not recommendations for a new setup. Groq and Stenox-hosted cloud services are not in the current selectable catalog. Choose a listed provider if an older profile is unavailable.
# AssemblyAI
Source: https://docs.stenox.app/providers/transcription/assemblyai
Configure AssemblyAI speech and understand streaming versus meeting requests.
Add an AssemblyAI key in **Settings → API Keys**, then select **AssemblyAI** in a speech profile. Choose its language and available region setting deliberately. Audio is sent to AssemblyAI.
## Models offered by this build
| Stenox label | ID |
| - | - |
| Universal-3.5 Pro Realtime | `universal-3-5-pro` |
| Universal Multilingual | `universal-streaming-multilingual` |
| Universal English | `universal-streaming-english` |
Streaming model selection applies to dictation. For files and Notetaker windows, the adapter uses **Universal-3.5 Pro**, with **Universal-2** as its configured fallback, and retains the chosen region and language. Do not infer a live streaming connection from a file or meeting request.
## Meeting behavior and privacy
Notetaker requests timed words and speaker labels. Labels are reconciled across requests and can be uncertain at boundaries. Correct the conversation when needed.
Stenox requests deletion of completed transcription jobs. Provider retention and account training settings still apply; this is not a zero-retention guarantee. Review [AssemblyAI's documentation](https://www.assemblyai.com/docs) and account controls.
## Account limits
Check current pricing, credits, quotas, and model access in the provider's console. Historical free-credit and per-minute amounts from older Stenox guides are not current promises. For failures, check the key, region, language, quota, and network connection before retrying.
## Diarized files
In **Transcribe file**, choose **Speakers & timestamps** to request a timed, speaker-labeled result from this provider. This mode preserves wording and disables cleanup/Polish; TXT, Markdown, JSON, SRT, and VTT exports are available. See [file transcription](/setup/file-transcription) for limits and recovery.
# Cohere Transcribe
Source: https://docs.stenox.app/providers/transcription/cohere-local
Run Cohere Transcribe locally with MLX and an explicit language selection.
Cohere Transcribe is a **local speech provider**, separate from cloud API services. Stenox uses MLX on Apple Silicon to run **Cohere Transcribe 03-2026 (8-bit)**.
| Model ID | Estimated download |
| - | - |
| `cohere-transcribe-03-2026-8bit` | \~2.4 GB |
The pinned model repository is `beshkenadze/cohere-transcribe-03-2026-mlx-8bit`. A first download needs internet; transcription runs on the Mac afterward. No Cohere API key is required for this provider.
## Choose a language
The adapter accepts Arabic, German, Greek, English, Spanish, French, Italian, Japanese, Korean, Dutch, Polish, Portuguese, Vietnamese, and Chinese.
Choose the spoken language explicitly. Automatic or unsupported language values normalize to English; they do not trigger multilingual auto-detection.
## Set it up
1. Edit Default or a custom profile in **Profiles**.
2. Select **Cohere Transcribe (Local)**.
3. Select a supported language and wait for the model to download and load.
4. Try a short recording before using a long file or meeting.
The model also appears in **Settings → Models**. [Model management](/setup/models) covers resource use and removal. Notetaker follows this shared speech selection and adds meeting-local speaker processing when needed.
# Deepgram
Source: https://docs.stenox.app/providers/transcription/deepgram
Use Deepgram cloud speech with explicit model and privacy choices.
Deepgram is an optional cloud speech provider. Add its key in **Settings → API Keys**, then choose **DeepGram** in a speech profile. Audio is sent to Deepgram and your account's limits and charges apply.
## Models offered by this build
| Stenox label | ID |
| - | - |
| Flux Multilingual | `flux-general-multi` |
| Flux English | `flux-general-en` |
| Nova 3 | `nova-3` |
| Nova 3 Medical | `nova-3-medical` |
Use a language supported by the selected model. The new default Deepgram choice is Nova 3; saved choices can differ. A specialized model name is not a guarantee of correctness for a particular recording.
## Dictation, files, and meetings
Flux is a streaming choice for dictation. File transcription and meeting audio windows use **Nova 3** when Flux is selected. Nova selections otherwise retain their model and language. Notetaker requests timed speaker labels and reconciles them across overlapping windows.
Stenox requests model-improvement opt-out (`mip_opt_out=true`). This is not a promise of zero retention. Check [Deepgram's documentation](https://developers.deepgram.com/docs) and your account settings for current service behavior, model access, and policy.
## If a request fails
Check the key, model access, quota, and connection. Optional cloud speech is not bundled with Stenox app access. Switch to a prepared [local speech provider](/providers/overview) yourself when needed.
## Diarized files
In **Transcribe file**, choose **Speakers & timestamps** to request a timed, speaker-labeled result from this provider. This mode preserves wording and disables cleanup/Polish; TXT, Markdown, JSON, SRT, and VTT exports are available. See [file transcription](/setup/file-transcription) for limits and recovery.
# Gemini Transcribe
Source: https://docs.stenox.app/providers/transcription/gemini-transcribe
Use Gemini speech with the same configured Gemini key and separate insight choices.
**Gemini Transcribe** is an optional cloud speech provider. It uses **Gemini 3.5 Transcribe** (`gemini-3.5-transcribe`) in the reviewed app build.
Add your Gemini key in **Settings → API Keys**, then choose **Gemini Transcribe** in a speech profile. This is the same key used by Gemini dictation Polish, but selecting one feature does not select the other.
## Dictation and meetings
Regular dictation uses the file speech adapter without requesting speaker labels. Notetaker adds its timed speaker request through that adapter. Audio is sent to Gemini; results arrive after processing the relevant audio window.
The adapter disables Interactions storage (`store=false`). Provider abuse-monitoring and account policies still apply, so this does not promise zero retention. Check [Gemini API documentation](https://ai.google.dev/gemini-api/docs) and your account for current access, limits, and billing.
The Gemini API speech integration is separate from the **Google Antigravity** subscription choice for meeting insights. Google Calendar context follows your selected insights provider independently of speech. Gemini API storage settings do not describe Antigravity's account policies.
## Diarized files
In **Transcribe file**, choose **Speakers & timestamps** to request a timed, speaker-labeled result from this provider. This mode preserves wording and disables cleanup/Polish; TXT, Markdown, JSON, SRT, and VTT exports are available. Gemini supports this mode for files up to 30 minutes. See [file transcription](/setup/file-transcription) for limits and recovery.
# OpenAI Transcribe
Source: https://docs.stenox.app/providers/transcription/openai-whisper
Configure the current OpenAI speech choices in Stenox.
The provider is now called **OpenAI Transcribe** in Stenox. This page retains its older URL so existing links still work; the current selectable models are not the old Whisper-only catalog.
## Configure speech
Add your OpenAI key under **Settings → API Keys**, then select **OpenAI Transcribe** in a speech profile. Choose a supported language and one of the options exposed by the app:
| Stenox label | App model ID | Use |
| - | - | - |
| GPT Transcribe | `gpt-transcribe` | Transcription after recording. |
| GPT Live Transcribe | `gpt-live-transcribe` | Realtime dictation. |
Files use **GPT Transcribe**, including when the live model is selected. Notetaker uses the adapter's file-transcription behavior and local speaker handling where required.
These are the names configured in the reviewed Stenox build. Your provider account must accept the requested model; the app catalog alone is not proof of account entitlement or successful inference.
## Data and billing
Audio is sent to OpenAI. API usage, quotas, and policies belong to your API account and are separate from a ChatGPT/Codex subscription and from Stenox app access. Check the [OpenAI platform](https://platform.openai.com) for account access and current service details.
OpenAI API keys in Stenox are for speech. The current dictation Polish providers are Local and Gemini. For Codex-backed meeting insights, use the separate [subscription integration](/meetings/processing).
# Parakeet
Source: https://docs.stenox.app/providers/transcription/parakeet-local
Local English and multilingual speech recognition through FluidAudio.
Parakeet runs locally through FluidAudio and Core ML on Apple Silicon. It needs a download before first use, then transcribes without sending audio to a speech service.
## Available models
| Stenox model | Model ID | Estimated download | Language |
| - | - | - | - |
| Parakeet v2 (English) | `parakeet-tdt-0.6b-v2` | \~2.5 GB | English |
| Parakeet v3 (Multi) | `parakeet-tdt-0.6b-v3` | \~2.5 GB | Multilingual |
**Parakeet v2 (English)** is the new-install speech default. Choose v3 for supported non-English speech. Do not assume that selecting a language makes the English-only v2 model multilingual.
## Set it up
1. Open **Profiles** and edit Default or a custom profile.
2. Choose **Parakeet (Local)** and the model you want.
3. Wait for download and loading to finish. Files already downloaded still need runtime preparation.
4. Record a short sample in the language you plan to use.
You can also pre-download models in **Settings → Models**. See [model management](/setup/models) for storage and loading behavior.
## Dictation and meetings
Dictation returns the transcript with enabled filler removal and user-defined replacements. AI rewriting is a separate **Polish** action. There is no automatic language-model cleanup pass merely because a cleanup model is selected.
Notetaker reuses the active Parakeet model and adds a local [speaker diarizer](/meetings/processing). It processes overlapping speech windows; words can arrive after a short delay. Overlapping speakers may be missed or assigned incorrectly, so review and correct important passages.
## Vocabulary
Use [Vocabulary](/setup/vocabulary) for explicit word or phrase replacements. The current Parakeet adapter does **not** send a boosted word list to its recognizer. Adding a replacement does not teach the model a name or guarantee that it recognizes previously unfamiliar speech.
## If recognition is poor
Check the selected model's language, input device, and microphone level. Try a short sample with less background noise. For meeting calls, use headphones to reduce duplicated system audio. [Whisper](/providers/transcription/whisperkit-local) and [Cohere Transcribe](/providers/transcription/cohere-local) are other local options; compare them on your own speech without assuming a universal accuracy ranking.
# Whisper
Source: https://docs.stenox.app/providers/transcription/whisperkit-local
Choose a local Whisper model through Stenox’s WhisperKit integration.
**Whisper (Local)** runs through WhisperKit on your Mac. The current Stenox app requires Apple Silicon and macOS 15 or later; selecting Whisper does not add Intel support.
## Available models
Sizes below are Stenox's catalog estimates, not runtime memory requirements.
| Model | ID | Estimated download |
| - | - | - |
| Tiny (English) | `tiny.en` | \~75 MB |
| Tiny (Multilingual) | `tiny` | \~75 MB |
| Base (English) | `base.en` | \~145 MB |
| Base (Multilingual) | `base` | \~145 MB |
| Small (English) | `small.en` | \~465 MB |
| Small (Multilingual) | `small` | \~465 MB |
| Medium (English) | `medium.en` | \~1.5 GB |
| Medium (Multilingual) | `medium` | \~1.5 GB |
| Turbo (Multilingual) | `large-v3-v20240930` | \~1.6 GB |
| Large v3 (Multilingual) | `large-v3` | \~3.1 GB |
| Distil Large v3 | `distil-large-v3` | \~1.5 GB |
The `.en` choices are English-only. Choose a multilingual model for other supported languages. Model size alone does not guarantee accuracy for your accent or recording.
## Select and prepare
Open **Profiles**, edit Default or a custom profile, and select **Whisper (Local)**. Choose a model and language, then wait for preparation. Alternatively, pre-download the model in **Settings → Models**. Existing profile choices remain in effect until you change them.
A smaller model needs less disk space and can be useful when comparing resource use. Test it with representative speech before relying on the output. [Manage local models](/setup/models) explains cached versus loaded state.
## Workflow limits
Whisper provides speech recognition. Word replacements and optional filler removal are separate steps; **Polish** is an explicit action after the transcript arrives. Notetaker can reuse the active Whisper model with meeting-local speaker processing. Review attribution and timing when the recognizer does not provide detailed word timings.
# Add provider API keys
Source: https://docs.stenox.app/setup/byok-configuration
Configure keys without confusing API usage, subscriptions, and local processing.
Open **Settings → API Keys** and add the key for the supported service you plan to use. Keys are stored in macOS Keychain. Do not paste a key into a profile prompt, Vocabulary entry, support message, or screenshot.
If an upgrade changed Keychain access, follow the [credential recovery steps](/guides/upgrading#recover-provider-and-license-access). Provider keys are separate from the Stenox license key entered in **Settings → License**.
| Service | Used by Stenox for |
| - | - |
| Deepgram | Cloud speech |
| AssemblyAI | Cloud speech |
| OpenAI | OpenAI Transcribe speech |
| Gemini | Gemini Transcribe speech and Gemini Polish |
## Select the service
A saved key does not switch your provider. Edit the speech profile for transcription, or enable **Polish on demand** in **Text cleanup** and choose Gemini for rewriting. Confirm the active profile before recording.
If your account rejects a request, check key validity, model access, quota, billing, and network connectivity in that provider's console. Stenox does not silently buy capacity or substitute another cloud provider. Switch to a prepared local model yourself when needed.
## API keys versus subscription sign-in
Codex, Claude Code, and Google Antigravity meeting insights use their subscription integrations in **Settings → Notetaker**. They do not use an OpenAI, Anthropic, or Gemini API key entered here. A consumer browser sign-in alone is not proof that the coding runtime is signed in.
See [cloud provider choices](/providers/cloud-byok), [meeting processing](/meetings/processing), and [local setup](/guides/privacy-first-setup).
# Transcribe a file
Source: https://docs.stenox.app/setup/file-transcription
Convert an audio or video file, optionally polish it, and save the result.
Open **Transcribe file** in the main window and drop a file or use the file picker. Select a speech provider, model, language, and output format, then start transcription.
The picker accepts common audio formats such as MP3, WAV, M4A, AAC, FLAC, OGG, WMA, AIFF, and CAF, and video formats including MP4, MOV, AVI, M4V, WebM, and MKV. Successful decoding also depends on the file's codec and the audio macOS can read; an accepted extension is not a guarantee that every file in that container will work.
## Choose processing
Local providers need their downloaded models. Cloud providers need a configured key and send audio to that provider. Filler removal follows **Text cleanup** settings. AI rewriting is offered only after the transcript is ready through **Polish**.
Deepgram Flux uses Nova 3 for file requests. AssemblyAI uses its file transcription path with Universal-3.5 Pro and Universal-2 fallback. A live streaming model label does not imply the file uses a streaming connection.
## Speakers and timestamps
Select **Speakers & timestamps** under Transcript when using **Deepgram**, **AssemblyAI**, or **Gemini Transcribe**. Other providers currently offer Plain text only; Notetaker's local diarization does not automatically add this capability to file transcription.
This mode preserves provider wording and file-relative timing. Filler removal, vocabulary cleanup, and **Polish** do not run on the timed transcript. Speaker labels belong to this file and can be unknown; they do not identify people by name.
Diarized files can be saved as **TXT, Markdown, JSON, SRT, or VTT**. Gemini's diarized-file path is limited to **30 minutes**. If the provider does not return usable speaker timings, the app reports an error rather than inventing them; retry or choose Plain text.
## Review, polish, and save
In **Plain text** mode, choose **TXT**, **Markdown**, or **JSON** for the result. Use Copy or Save after reviewing it. If you request **Polish**, **Undo Polish** can return to the pre-polish result. Prepare the cleanup model before using this optional action.
The file workflow is separate from the meeting transcript menu. For a captured session, use [Notetaker exports](/meetings/exports).
If decoding fails, try a readable WAV or M4A copy. If a provider rejects the file, check its current limits and account status. An active meeting reserves the shared speech runtime, so finish it before attempting a competing job.
# Keyboard shortcuts
Source: https://docs.stenox.app/setup/hotkeys
Configure dictation and meeting shortcuts and understand their actions.
Open **Settings → General** to review or change your shortcuts. Saved choices take precedence over these new-install defaults.
| Action | Default | Use |
| - | - | - |
| Hold to dictate | Fn / Globe | Hold while speaking, release to finish. |
| Toggle dictation | Option–Space | Press to start, press again to finish. |
| Cancel recording | Escape | Cancel the current dictation. |
| Undo dictation | Option–Z | Undo when Stenox has a valid insertion or clipboard context. |
| Start / end meeting | Control–Option–M | Start a Notetaker session or end the active one. |
The meeting shortcut can start directly without the setup sheet. A meeting and dictation cannot record simultaneously.
## If a shortcut does not work
Check the saved chord, conflicts with macOS or another app, and whether Stenox is running. Some external keyboards do not expose Fn/Globe as expected; choose another available shortcut. Confirm that speech preparation and required permissions are complete.
Undo depends on the target and whether Stenox can verify the text it inserted. It is not a general undo command for arbitrary changes in another app. Use the target app's own undo when appropriate.
The current build does not document the old right-Command/right-Option feedback shortcut as an available workflow. [Send feedback](/support/feedback) through the supported contact options.
# Manage local models
Source: https://docs.stenox.app/setup/models
Download, select, load, and remove local Stenox models.
## Download and select
Open **Settings → Models** to browse local model files and download them. Downloading a model in the background does not switch the active selection. Select speech in your [profile](/profiles/creating-profiles), and choose a language model in **Text cleanup → Local** for optional Polish.
Selecting an uncached local speech model starts its download and preparation. A cached model loads from disk. Initial downloads need a network connection; inference then runs on your Mac.
| State | What it means |
| - | - |
| Not downloaded | Model files are not ready on disk. |
| Downloading | Stenox is fetching the files; keep network access and storage available. |
| Downloaded / cached | Files exist locally, but the model may not be in memory. |
| Loading | Stenox is preparing the model for inference. |
| Ready | The selected runtime is loaded and can process input. |
Download labels are estimates from Stenox's catalog. Actual storage includes supporting files, and runtime memory can exceed download size. Running local speech, a meeting diarizer, and meeting insights together uses more memory than speech alone. No fixed per-model RAM minimum is guaranteed here.
## Loading and memory
Stenox prepares the selected local speech model on startup and keeps eligible selected models ready across ordinary idle periods. It restores eligible cached models when needed; changing a selection or disabling local cleanup releases the old runtime through the model lifecycle. Do not expect a universal five-minute unload timer.
Notetaker borrows the active speech model. While capture holds that model, competing jobs and model changes are guarded. Pausing retains meeting state. Ending releases the reservation and meeting-only models while keeping dictation's speech selection ready.
**Polish** waits for a selected model that is already warming up, but does not silently download a missing cleanup model or use another provider. A cleanup failure leaves the dictated text available.
## Remove downloaded files
Use the delete control beside a downloaded model in **Settings → Models**. Read the confirmation, including any profiles that use it. Removal is guarded while a model is loading or reserved. Removing files means they must be downloaded again before use; unloading a runtime only releases memory.
## Catalogs
* [Parakeet](/providers/transcription/parakeet-local) and [Whisper](/providers/transcription/whisperkit-local)
* [Cohere Transcribe](/providers/transcription/cohere-local)
* [Local AI cleanup and legacy models](/providers/llm/mlx-local)
* [Meeting-only processing models](/meetings/processing)
# Word replacements
Source: https://docs.stenox.app/setup/vocabulary
Set explicit word and phrase replacements for transcribed text.
Open **Vocabulary** in Stenox to define what a recognized word or phrase should become.
Put the text you want to replace in **Word or Phrase**, for example `postgress`.
Put the desired output in **Replace With**, for example `PostgreSQL`.
Add the entry and dictate a short example. Check the actual recognized text if the replacement does not apply.
## What replacements do
Stenox applies explicit replacements as part of transcript cleanup. They work independently of the optional **Polish** action. A blank replacement keeps the original rather than defining new text to insert.
This is not automatic vocabulary learning, inference of unfamiliar names, or recovery of misheard words from screen context. The current Parakeet adapter does not implement recognizer vocabulary boosting. Some cloud adapters can additionally send supported vocabulary hints, but that does not turn every local model into a trainable dictionary.
Keep entries narrow and check a sample after changing them. If the recognizer produces a different phrase from the one you entered, an exact intended replacement may not be triggered. Review important names and identifiers before using the result.
# Send feedback
Source: https://docs.stenox.app/support/feedback
Report a reproducible issue without sharing keys or private conversation data.
Email [support@stenox.app](mailto:support@stenox.app) with a short description of the issue. The current build does not provide the old automatic voice-feedback workflow described in earlier docs.
Useful details include:
* Stenox version from **About**, macOS version, and Mac chip.
* The screen or feature involved and the selected provider/model.
* Steps to reproduce, what you expected, and what actually happened.
* The visible error and whether a short non-sensitive example reproduces it.
If a screenshot helps, remove API keys, account details, and private transcript or reference content first. Do not send your session database or recordings unless you deliberately choose to share relevant material with support.
For a missing feature, state whether you are using the public download or a development build. Include the version shown in **About Stenox** and check [what’s new](/changelog).
# Get help
Source: https://docs.stenox.app/support/overview
Find setup guidance or contact Stenox support.
Start with [Troubleshooting](/guides/troubleshooting) for permission, licensing, model, dictation, or meeting issues. Read [Upgrading Stenox](/guides/upgrading) for upgrade preparation and credential recovery. Check [Release and access](/pricing/overview) if your build does not have a feature shown in these docs.
Email [support@stenox.app](mailto:support@stenox.app) or visit the [support page](https://stenox.app/support). Include the app and macOS versions, the affected feature, selected provider/model, and what happened.
For third-party limits or billing, also check that provider's console. Stenox app access and external provider usage are separate.
[Send useful feedback](/support/feedback) explains what to include and what to leave out.