1AudioTool v0.3.0 - Live translation, on local machine

Added

• Live translation, on this machine — talk in one language and the captions
  arrive in another, about a second behind your voice. Each sentence is
  translated during the pause that follows it, so nobody waits for you to stop.
  • Settings - Speech picks the caption language, and the pack for that pair is
    a download with its size and its progress. Nothing ships in the bundle and
    nothing downloads on its own: until you take a pack, the feature says it
    needs one rather than fetching it behind you.
  • Captions window opens the same captions at reading-across-a-room size —
    three type sizes, a high-contrast switch, and it stays above other windows.
  • Captions on this network hands anyone on the same Wi-Fi a read-only
    page, as a link to copy or a QR code to scan on a phone. It carries the
    pairing token and dies with the pairing session after fifteen minutes; the
    page then stops updating and says so.
  • Keep the sentence you spoke above its translation, or show only the
    language your audience reads.
  • The transcript stays what was heard. A translation is never pasted, never
    stored as the transcript, and never trusted to carry names or filenames.
• A shelf of speech models, with a Settings tab of its own — Settings -
  Models lists every pack the app can use, with its size, a line on what that
  size buys, a Download button and progress. A fresh install shows the same
  shelf's door on Dictate instead of a picker that only says "No model
  installed", and the first pack to land becomes the profile's model, so the
  next dictation just works.
• An update is now something you read before you take it. When 1AudioTool
  finds a newer signed release it opens a window with the version, when it was
  published, and the full release notes — headings, bullets, emphasis and any
  screenshots the release carried — then offers one button that downloads it and
  restarts. The download reports its size as it goes, and if anything fails it
  says so rather than leaving the button looking pressed.
  • Later asks again next time you open the app; Skip this version stops
    that one release being offered at all. Either way Settings - General still
    has it, and asking for it there is what un-skips it.
  • The check runs shortly after launch instead of only when you go looking, so
    the status line can say an update is waiting. On a first run it waits its
    turn behind the question about anonymous counts.
• Open a meeting — a Listen recording now has a page of its own. The
  recording plays back and every speaker turn seeks it, the transcript can be
  made again from that recording with today's languages and a bigger model, and
  correcting it there teaches the same phrase rules a corrected dictation does.
  A second pass rebuilds everything the first one produced: speakers, chapters
  and action items.
  • The language picker is in Listen as well as in Settings, because the
    languages a meeting is decoded with are a decision made where the meeting
    starts.
  • A note says when the languages and the model disagree — an English-only
    pack cannot mix languages — and a transcript that came back mostly silence
    says so plainly rather than leaving you to infer it.
• Play back the clip you just dictated — the result card has a player with a
  scrubber you can seek with the arrow keys. A clip whose recording was not kept
  says so instead of showing dead controls.
• Read your corrections with a coding CLI you already have — Review's new
  Analysis tab groups the review queue by cause: the vocabulary the model keeps
  missing, the homophones, the places a second language starts mid-sentence. It
  names the near-duplicate rules that should be one rule, and says what to do
  about each group. Settings - Agents lists the coding CLIs found on this
  machine and which one to use; the default is to ask each time, and nothing
  runs until you press the button. This is the one feature that sends anything
  off this machine — see Privacy.
• The room says when it is working against you — a dictation now notices
  music or a TV playing behind you, a room your voice comes back off, and a
  background sitting nearly as loud as you are. It names the worst one, once,
  quietly, and never stops the recording. Settings - Speech turns it off.
• Every paste mode explains itself — Verbatim, Clean, Prompt, Message and
  Command each show one line on what lands in the clipboard and the same spoken
  sentence run through them, so a mode can be understood without spending a
  recording to find out.

Improved

• The Inbox is now Review — the same clips waiting to be looked at and the
  same rules waiting to be learned, with the Analysis tab beside them.
• Closing the main window no longer stops dictation working. It hides, the
  tray brings it back, and in between the shortcut still starts and stops
  recordings. Quit still quits.
• The app stops drawing what nobody is looking at — a hidden window stops
  rendering instead of painting to a screen that is not there, a level meter
  releases the microphone when its panel is put away, and the timers in a window
  you cannot see stop counting. This was background processor use, on both
  platforms, for doing nothing.
• The shortcut is written as an instruction — "Double-tap ⌥ to stop" rather
  than a pair of symbols you have to decode.
• Actions is hidden while the action engine is unfinished — it was a space
  you could open and a mode you could choose that did not yet do anything for
  you. Command mode is still in the mode list, now marked as not yet wired up,
  and delivers Verbatim until it is.

Fixed

• Fixed every recording in a signed release failing with the microphone
  blocked, while System Settings showed 1AudioTool switched on — the shipped
  build did not declare the microphone the way macOS demands of a notarized app,
  so the toggle in Privacy & Security granted nothing and no prompt ever
  appeared. Release builds now declare it, and a release is checked for that
  declaration before it ships.
• Fixed the double tap not stopping a recording when permission arrived after
  launch — watching for the gesture needs Accessibility or Input Monitoring,
  and granting it used to take a restart before the gesture came alive. The
  watcher now starts within seconds of the grant, and a recording begun from
  inside the app can be stopped the way the overlay says it can.
• Speech models can now be downloaded from inside the app, so a first
  recording on a new install no longer fails with "whisper-cli was not found".
  The transcription path looked for an engine before it looked for a model, and
  reported a missing model as a missing developer tool. It now says that no
  model is installed and where to get one — and when the bundled engine has not
  finished loading the model, it waits for it rather than falling back to a
  command-line tool that packaged builds do not carry.
• Fixed Vietnamese input methods (UniKey, EVKey) stopping while 1AudioTool
  runs on Windows, and Control shortcuts arriving out of order. The
  double-Control watcher swallowed every Control press for a quarter of a
  second and re-injected it later, which starved other keyboard hooks and
  reordered chords. It now only listens, like the Option watcher on macOS:
  every key passes straight through, nothing synthetic is posted, and an input
  method's own replacement text is left alone.
• Fixed Esc doing nothing on the overlay once it had been clicked on
  Windows. The panel now answers Escape itself — cancel while listening or
  transcribing, dismiss on a failure — in addition to the global watcher.
• Fixed a black console window flashing on Windows whenever the app started
  its speech or translation engine, downloaded a model, checked the licence,
  or read the machine id. Every helper the app runs is now started without a
  console.
• The Windows helpers are built against the static C runtime, so they start on a
  machine that has no Visual C++ redistributable installed.
• Fixed pressing Install doing nothing for several seconds, and sometimes
  failing after the app had already offered the update. Installing re-ran the
  whole check over the network before it began, so the press could not be
  answered until a second round trip had finished, and that round trip had its
  own way to fail. The release found by the check is now the one that installs.
• Fixed a recording cancelled while the microphone was still opening leaving
  the microphone open, and stopping one recording muting the next when the
  two overlapped. A capture that is no longer wanted now releases its input the
  moment it arrives, and only the capture that owns the audio is allowed to
  tear it down.
• Fixed a correction you saved looking as though it had been thrown away —
  the result card went back to showing what the model heard rather than what you
  corrected it to.
• Fixed Prompt mode cutting your sentence at the dot in a filename — "fix
  the login bug in auth.ts before the demo" arrived as "fix the login bug in
  auth." A full stop now has to end a sentence before it is treated as one.

Privacy

• Live translation runs entirely on this machine. Translation packs are
  downloads that then work offline, and no sentence is ever sent anywhere to be
  translated.
• The captions page on your network is read-only, needs the pairing token, and
  expires with the pairing session after fifteen minutes. It is off until you
  turn it on, and it serves only the captions of the session you are running.
• Transcript analysis is the one thing in 1AudioTool that leaves this
  machine, and only when you press the button. The run hands y...

1AudioTool v0.2.0

Added

• Speaker separation in meetings — a Listen recording now shows who said
  what. Voices are told apart on this Mac after the meeting stops, each turn is
  attributed, and clicking a name renames that speaker everywhere in the
  transcript.
  • Teach my voice records a few seconds of you and labels you Me
    instead of a number. Nothing is ever labelled Me until you enrol, and the
    voiceprint never leaves this Mac.
  • Two engines, chosen per profile in Settings - Speech. Built-in needs
    nothing extra and separates an ordinary two- or three-person meeting.
    Speaker model is a download that is sharper on noisy calls and on
    voices that sound alike.
  • Markers you drop during a meeting and the chapters built from them appear
    beside the speaker list, so a long recording can be read before it is
    replayed.
• Pick a microphone and see it working — the input picker in Dictate,
  Settings and the recording overlay lists every microphone by name and type,
  Built-in, Bluetooth, USB, Virtual or External, and draws a live level meter
  under the list. A microphone that another app is holding, or that has quietly
  stopped responding, is now visible before a recording is spent finding out.
  • The picker is on the overlay while a dictation runs, so you can move to
    another microphone without losing what you have already said.
  • When a capture fails, the same picker is on the failure card, and choosing
    an input there is what retries it.
• Create a profile from the sidebar — choosing "New profile" in the profile
  menu opens a name field in place. The new profile starts on the default
  settings and becomes active straight away, so a second voice or a different
  set of languages is a few keystrokes away. The free tier keeps one profile,
  and the menu explains what Pro adds rather than failing silently.
• Long dictations no longer make you wait at the end — each sentence is
  transcribed during the pause that follows it, so when you stop, only the last
  one is left. On a recording of about a minute and a quarter the wait after
  stopping went from 2.7 seconds to a third of a second, and it no longer grows
  with how long you spoke. With live text turned on, each sentence appears in
  the overlay as it lands. It needs the quality model loaded, and it can be
  turned off in Settings - Performance - Transcribe as you speak.
• Dictate again from the result card — the card that shows a finished
  transcript starts the next dictation, without reaching for the shortcut.
  Double-clicking the transcript there opens it for correction.
• Discard a recording from the tray — Escape only reaches a window that has
  focus, which is never the case when you dictate into another app.
• Windows support — the double-Control gesture, insertion at the caret in
  the app you were last in, spoken responses through the system voice, and the
  folder picker all run natively.
  • Releases now carry a Windows installer alongside the macOS bundles, with
    the same update signature. It is not Authenticode-signed yet, so SmartScreen
    warns on first run.
• The Inbox pages — Needs review, Learning, saved rules and Sources each
  show a page at a time with Previous and Next. A full Inbox used to build a
  transcript editor and an audio player for all 500 clips before it would draw.

Improved

• The global gesture on macOS is Option twice, not Control twice. macOS
  binds double-Control to its own Dictation and keeps that binding even when
  Dictation is switched off, so reclaiming it meant swallowing every Control
  press system-wide and posting a synthetic one back a fraction of a second
  later. Nothing binds a bare double tap of Option, so the watcher now only
  listens: no modifier is ever swallowed or replayed, and Option-click,
  Option-drag and every Option chord are untouched. Existing profiles move
  across on their own. Windows keeps Control twice, where a lone Alt press
  would open the menu bar.
• Dictation is faster to finish — the model now loads while you are still
  speaking rather than after you stop, and it stays loaded between dictations,
  so the wait between finishing a sentence and seeing it appear went from about
  a second to about a fifth of one on the machine this was measured on. The
  timing, and how long the model stays loaded, are in Settings - Performance.
• The transcript reaches your cursor sooner — pasting used to wait for the
  history entry, the recording and the profile to be written first; now it goes
  first and the bookkeeping follows. Bringing your app back to the front no
  longer waits out a fixed pause, and no longer asks macOS for permission to
  control other applications.
• A profile that names a model you do not have now corrects itself — it
  used to surface, after a whole recording was spoken, as a microphone error.
  The choice is matched against what is actually installed and written back, so
  Settings shows the model that is really being used.
• Copying on Windows keeps every character — text outside the system's
  console character set, accents, Vietnamese and emoji, now copies intact, and
  pasting no longer waits out a fixed pause before it starts.
• Saving after a clip no longer rewrites your whole history — transcripts
  are stored separately from the profile, so the pause some people saw a moment
  after pasting is gone. Existing history moves across the first time this
  version opens it.
• The Inbox holds less memory — recordings are read only when their row
  scrolls into view, and a list releases the recordings the previous render
  held. Re-rendering the Inbox used to leak one copy of every clip's audio.
• Settings are written once per pause rather than once per change.
• The interface is rebuilt, and everything carries over — profiles,
  transcripts, correction rules, packs, appearance and your licence are
  untouched, and the app reads exactly the same files on disk it did before.
  Each space now loads when you open it, so the window is ready sooner and
  stays lighter.
• Supporting text in the rail and panels is a shade darker — it was just
  under the contrast a small grey label needs to be comfortably readable.

Fixed

• Fixed dictation from the shortcut always failing when the app was not on
  screen — minimised, hidden behind another app, or on another Space, the
  shortcut would show the overlay with no timer, no microphone name and no
  input picker, then fail with "did not respond, on two attempts". The system
  will not open a microphone for a window that is not visible, and it does not
  refuse the request either; it simply holds it until the window comes back,
  which for a dictation into another app never happens. The recording overlay,
  which is on screen for exactly as long as a capture runs, now opens the
  microphone. Dictating from another app works the same way it does inside the
  app.
• Fixed the end of the last word being cut off — the final fraction of a
  second of every recording was dropped as the microphone was released. It is
  now waited for and included in the clip.
• Fixed the overlay going blank after the main window was reloaded — it
  stopped accepting updates for the rest of the session and showed a stale
  card.
• Fixed a shortcut that never armed saying nothing — watching for the
  gesture needs Input Monitoring, a separate grant from Accessibility, and
  without it macOS refuses silently while the settings panel still shows the
  shortcut as bound. 1AudioTool now asks for that permission, reports it
  alongside Accessibility, and links straight to the right pane when it is
  missing.
• Fixed the app not responding after a dictation — pasting at the cursor
  ran the clipboard write, the app switch and a deliberate settle delay while
  the window was trying to draw, and working out which app was frontmost added
  two more round trips before every capture.
• Fixed downloading a model freezing the app — the window stopped drawing
  and macOS showed "Application Not Responding" until the model finished,
  several minutes on the larger ones. The progress bar now moves while it runs.
• Fixed activating or validating a licence, choosing a project folder,
  scanning a folder, reading storage usage and listing voices blocking the
  window while they wait.
• Fixed saving a recording stalling the window — each clip crossed to the
  backend as a list of a few million numbers, which the app built and parsed
  itself; it is now sent as raw audio.
• Fixed typing in a search box, and the level meter while recording,
  competing with the interface for frames — searching redraws once you pause,
  and finishing a clip prepares the audio out of the way.

Privacy

• Speaker separation runs entirely on this Mac. Voice vectors and the
  voiceprint from "Teach my voice" are computed and stored locally, are never
  uploaded, and no speaker is labelled Me until you explicitly enrol.
• Audio, transcripts, profiles, correction memory and licence state still stay
  on this machine unless you export them. There is no account, no API key and
  no server.
• The speaker model is an optional download, like the transcription models. It
  runs on this disk once installed, and meetings are never sent anywhere to be
  separated.

1AudioTool v0.1.0

Added

• Fully local macOS dictation powered by whisper.cpp, with Verbatim and Clean output modes.
• Speaker profiles with per-profile languages, model selection, vocabulary, microphone choice,
  correction memory, and transcript history.
• A global Control-twice shortcut, floating recording overlay, clipboard delivery, and optional
  insertion into the previously active application.
• Recheck reports that separate genuine model improvement from correction-rule improvement.
• Free and Pro licensing with local entitlement verification and privacy-preserving storage.
• Signed in-app software updates delivered from the dedicated public releases repository.
• Nine spaces for the whole voice loop — Dictate, Inbox, Memory, Context, Actions, Listen,
  Insights, Studio, and Settings.
  • Inbox collects clips that still need review and spelling rules waiting to be learned.
  • Memory searches, filters and pages every transcript, with audio replay while correcting.
  • Ask memory answers a question from what you have dictated and cites the transcript it came from.
• Speak more than plain text — Prompt, Message and Command modes reshape what you said into a
  well-formed prompt, a written message, or a parsed action, locally, before it reaches the clipboard.
  • Voice actions always show a preview, and anything consequential needs a keypress to confirm.
    A matching voice is never authorisation.
• Personal dictionary and project context packs — bias transcription toward your own
  identifiers, filenames and jargon. Point a pack at a repository and it learns that project's
  vocabulary, so readNativePromptBuffer comes out spelled right the first time.
• Correction memory that compounds — correct a transcript once and the phrase is remembered for
  that speaker, with rules you can scope, merge, edit and forget.
• Insights — how much you have dictated, which applications and languages you use it in, and a
  streak that survives history retention because the daily rollup is kept separately from transcripts.
• Listen — turn a day of captures into a written digest you can read first, then have spoken back.
• Phone as a microphone — pair a phone over the local network with a QR code and dictate into
  this Mac from across the room.
• Microphone pinning — pin the input device so a headset or phone connecting mid-session cannot
  quietly move the system default out from under a recording.
• Personal accuracy benchmark — read a script aloud and measure word error rate, project-term
  recall and latency on your own voice, then pick a model from what was actually measured.
• Model downloads in the app — fetch base.en (141 MB), small (466 MB) or large-v3-turbo
  (1.6 GB) with progress, and switch between them per profile.
• Mixed-language speech — select up to three spoken languages for one profile, for people who
  code-switch mid-sentence.
• Live words on the overlay — watch the transcript build while you are still speaking instead of
  waiting for the clip to finish.
• Backup, restore and profile transfer — take everything this install holds to another machine.
  Profiles export as JSON, or as an AES-GCM encrypted file when you want one.

Improved

• Accuracy defaults that match how people actually dictate — new profiles start on explicit
  English rather than auto-detect, because a short clip in accented English is easily classified as
  the speaker's first language. The app warns when a short recording still uses Auto-detect.
• Cleaner audio before transcription — recordings have DC offset removed, outside silence
  trimmed and conservative gain applied, and the app reports quiet, clipped or mostly-silent input
  instead of leaving it to look like a model failure.
• Speech enhancement is now a choice — browser echo cancellation, noise suppression and
  automatic gain can smear consonants on some microphones, so each profile can A/B test it against
  the raw microphone.

Privacy

• Audio, transcripts, profiles, correction memory, and licence state remain on this Mac unless the
  user explicitly exports data.
• There is no account, no API key and no server. Transcription runs against a model on this disk.
• Backups deliberately exclude microphone recordings and the licence key.