最近更新

The first public version, a beta: voice input for macOS on remote models, from actnow.dev. It updates to the next beta and to the release that follows (Settings → About → Beta updates). 0.1.0-beta.1 was built but not released; its changes are listed here.

Added

• Transcribe: a shortcut (fn by default; tap, double tap or long press, several per feature) records, a speech service recognises, a language model cleans up (fillers, false starts, punctuation, numbers; an optional structure layer and your own rules), and the text is pasted where the cursor is. Without a language model the text goes in as recognised; when it fails or is slow, too.
• Translate: the selection (read through Accessibility, or ⌘C where that is safe) or what you say, into the language you choose; a strict retry when the answer looks untranslated; an answer card to read, copy, insert or translate again into another language.
• Rewrite: say how to change the selection, or what to write with nothing selected; every answer is written one line under the pill, then waits in the answer card (as in Voxt), where Insert puts it in place (over the selection) and Continue records a follow-up, answered in plain text in the card's conversation.
• App enhancement: groups of rules by app and by website (address rules read through Accessibility, no browser automation), a scene prompt layered on top of the clean-up, the key to press after inserting, the features each group applies to, a live match preview, and the group's label in the overlay and the history.
• Dictionary: terms with aliases, sent as hints to the speech services and as a glossary to the language model; CSV and text import and export.
• Model services: Doubao and Alibaba Cloud Model Studio streaming recognition and any OpenAI-compatible transcription API; DeepSeek, Qwen, Volcengine Ark, Kimi, GLM, MiniMax, StepFun, MiMo, SiliconFlow, OpenAI, Anthropic, Gemini, OpenRouter and other OpenAI-compatible language models; keys in the Keychain, connection tests, a service per feature.
• Home: a slogan with your Transcribe shortcut in a key cap; what is still missing (a service, a permission, the shortcut), each line the way to fix it, and nothing once all is ready; a try of the whole chain on the page until the first result; today's results and characters, the last five results.
• Setup guide on the first start (and from Settings → General): privacy, permissions, services, then a real try of Transcribe, Translate and Rewrite, and App enhancement and the dictionary.
• History: by feature, searchable, with the texts, where the result went, the services, the rule that matched and each stage's time; recordings kept on request. Right-click an entry, in History or among Home's recent results, to copy its result or its recognised text, show its details, show its recording in Finder, retry it or delete it (with its recording, after a confirmation); a selected History row takes ⌘C and ⌫ too.
• The recording overlay (non-activating, so the app you type into keeps its focus), after Voxt's: the feature's icon and a waveform on a translucent, blurred pill with no buttons and Voxt's proportions (the shortcut stops, Escape cancels; the pointer goes through it), one line of live text under it, the pill as wide as the text up to 300 and the text then scrolled to its newest words, a loader while it works, errors in place of the waveform, nothing after an insert; and the answer card, with start and stop sounds. Every session starts as the pill, and it grows about its edge of the screen without moving; every flow shows the same pill until its result (what the model works on, then its draft, on one line), and the card comes only with the result.
• The answer card, after Voxt's: 440 wide (Voxt's is 560) and as high as its text (up to 220 of it, then it scrolls), its actions in the header (the target language as a capsule opening our own menu in a window of its own, Continue, Insert, Copy, Close), the header as far from the top as from the rule, Voxt's own icons in the pill and the card, why the result is in the card in Insert's tip instead of a line of orange, Copy marking itself for two seconds, Insert hiding the card while it pastes, another language translated in place (the old text with a spinner until the new one comes), and Rewrite follow-ups recorded in the card as a conversation (a small waveform, bubbles); Escape closes it, or cancels a follow-up only.
• Silence left out before recognition (a light energy detector, no model): a file service gets the recording without the silence around the speech and with long pauses shortened, a streaming service nothing before the speech starts, and a recording without speech is not sent at all (Settings → Audio → Skip silence).
• Sound optimization before recognition (pure DSP, no model): a high-pass filter at 80 Hz takes out hum and rumble, and one gain measured on the speech alone raises a quiet voice (up to +12 dB, never lifting the room's noise over -50 dBFS) or lowers a very loud one a little, with a limiter at -1 dBFS; a stream's gain follows slowly and holds through pauses. The recording kept in the history is the original; the history says what was done and the gain (Settings → Audio → Optimize sound, on by default).
• A small window (800×520 by default, 680×440 at the least) with no status bar: notifications stack as cards at the foot of the sidebar (a dot on Settings when it shows only icons), and a missing service or a stopped helper becomes a card with the action that fixes it.
• The menu bar icon: start or stop transcribing, paste the last result, the last three results to copy (one short line each, the whole in the tooltip); it changes while recording. Closing the window keeps Auralink there, and it can open at login.
• Settings: General (languages, the menu bar, login, the setup guide), Permissions, Audio (the microphones in your order of preference, skipping silence, optimizing the sound, sounds), Overlay style (place, opacity, corners, the recording time, with a live preview over a desktop picture), Shortcuts (every feature's keys and their conflicts), Appearance (the sidebar shrinks to icons in a narrow window), Network (proxy), Privacy, Storage, About (versions, updates, the self-check, reporting a problem); English and Simplified Chinese.
• Signed with a Developer ID and notarised. Updates from actnow.dev, in Auralink's own window: a daily check that only shows "Update available" at the foot of the sidebar (a dot on the icon rail), a panel with the new version's notes in Auralink's language (English where a version has no Chinese, marked so), a download of only what changed with its progress, installed when you say so, then Restart to Finish the Update; a failed download leaves Auralink unchanged and offers Try again. Check for Updates… in the app menu, the menu bar icon's menu, the palette and Settings → About, which also shows the changelog of every version in Auralink's language. An update installed while Auralink runs asks for a restart instead of starting the new helper.
• Beta updates (Settings → About): betas too, or releases only; on by default in a beta. Turning it off never goes back to an older version: a beta stays until a release newer than it. Install updates automatically (off by default) installs what the daily check finds and asks for the restart.
• Mute system audio while recording (Settings → Audio, on by default): the current output device is muted from just after the start sound until the recording stops, so that videos, music or calls are not recorded; a device you muted stays muted, one you unmute during the recording stays on, and a device that cannot be muted leaves the recording as it is (the overlay says so). No permission is needed. If Auralink or its helper stops while muting, the sound comes back at once, or at the next start.
• Sounds to choose (Settings → Audio → Play sounds): the start and the stop sound are each chosen from the ActNow family's sound library (from React Bits Sound Lab: taps, toggles, media, system sounds…; Tally and Wrap by default), or macOS's system sounds, or none. A compact picker searches them (names and tags), filters by type, plays each with ▶ (again to stop) and sets their volume in steps; a Try button plays the choice. A Voxt pair chosen in an earlier build is kept as those system sounds. The output is muted once the start sound has been heard, not after its file's silent end. The family's sounds are in the builds from actnow.dev; a build from the source code plays macOS's system sounds.
• History settings (the button beside History's search field): keep the recordings (the same switch as Settings → Privacy), choose the folder they are kept in (the recordings already kept move there, each one copied and checked before the original is deleted; if the folder is on a drive that is not connected, new recordings go to the default folder until it is back), see how much room they take, export them (all, or those of the tab shown) into a folder you choose, named by their date and first words, or delete them all and keep the texts. Automatic clean-up, on as in Voxt: entries older than 3 months are deleted with their recordings; you can choose 1 week to 1 year, keep at most 50 to 1000 entries of each feature, or turn it off. A new setting that would delete entries now asks first, with their number.
• Retry a history entry from its recording: a failed entry (or one with no text) shows Retry on its row instead of Failed, in its details and in its context menu; other entries with a recording have Transcribe again. The recording is recognised again (streaming services get it through a stream, faster than it was said), then cleaned up, translated or rewritten with your settings as they are now; the entry is updated in place (text, status, time), and nothing is pasted. The overlay's notice of a failed recognition opens the entry and retries it.
• When recognition fails, times out or finds no text, the entry keeps the recording for 7 days even when Keep the recordings is off, so that you can retry it; a successful retry deletes it.
• Logs (the Logs page and panel): click a line to copy it; Shift-click, ⌘-click or drag to select lines as in Finder, ⌘A for every line shown; right-click to copy a line, the selection or every line shown. The header shows how many lines are selected, with Copy and Deselect.

Changed

• The note of which output device Auralink muted is kept in Application Support instead of ~/Library/Caches, so that a cache cleaner run after a crash cannot leave your speakers muted. Logs leave out keys, tokens and proxy passwords.
• Auralink is free and open source, under the GNU AGPL-3.0: the source is on GitHub, and contributions are welcome (CONTRIBUTING.md; a contributor licence agreement). Versions up to 0.1.0-beta.1 were under the Apache License 2.0. Settings → About shows the copyright and the licence, links to the source code, the website and the changelog, and opens the licence and the third-party notices, which now ship inside the app. Report a bug opens a GitHub issue with the diagnostics filled in.
• Downloads, updates and release notes come from actnow.dev instead of GitHub Releases: sign in at actnow.dev/download/auralink to download. Updates come from actnow.dev/updates/auralink, and Beta updates read its beta channel, which always offers the newer of the latest beta and the latest release.
• The menu bar icon is now a speech bubble with a waveform, in thin lines like the system's own icons; while recording, the bubble turns solid.
• A result goes where you are typing when it is ready, not only into the app you started in (as in Voxt): move to another app and click into a text field while Auralink records or works, and it is pasted there; web editors such as Notion or Slack in the browser count as text fields. The app's group's key after the paste is pressed only in the app you started in, whose group shaped the result, and a selection is replaced only in the app it was selected in. Where no text field is focused (Finder, the desktop, a page you are reading) or the focus is in a password field, the result waits in the answer card, which says why; click where it should go, then Insert. Clicking Auralink's own window still sends it back to the app you started in. The history shows the app it was inserted into when that is another one. Insert in the answer card and Paste the last result follow the same rules.

Known limitations

• Apple silicon and macOS 15 or later only (no Intel build yet).