changelog.

a fix for macs that said no sound from MacBook Pro Microphone on every press.

on some macs the speakers and the mic run at different rates, like 44.1 kHz and 48 kHz. since 0.9.4 the app listened to the mic at the speakers' rate, got nothing, and said there was no sound. now it listens at the mic's own rate, whatever the speakers are set to. if this hit you, nothing else changes: update and press.

if the mic's rate changes while the app is open, the next press is heard too. before, that one press came back empty.

everything else is 0.13.0.

any model for either job, and a tiny new one.

dictation and meetings pick from the same five models now. before, dictation had the two parakeets, and meetings had whisper and parakeet v3. now whisper can take your dictation if you don't mind waiting for it, and parakeet v2 can hear a meeting in english. each card says what the model costs in the job you're picking for. a model you downloaded for one job is already there for the other.

the new one is whistle, from cactus compute, out since october 2. it is a 17 mb download where the others are hundreds of mb or more, and it runs on the cpu. it hears english, german, french, spanish, italian, dutch and polish. its card says experimental: it is new, and on meeting audio whisper is better. try it and see.

whisper and whistle read about 30 seconds at a time, so a longer dictation is cut where you pause, and no words are lost at the cut. a slow model gets more time before the app decides it is stuck.

settings can remove any model's download, not only parakeet's. picking whisper for dictation does not set up meetings, so the app won't start asking about your calls.

copy diagnostics now says which model answered each press. it never did.

0.12.0 never shipped on its own, so its two lines in the app arrive with this one.

a new site, a readme, and two lines in the app.

the site says what this is for now: talking to your agents. you can hold a key on it and watch a prompt land in an agent's box, after the same wait the app takes. you can press record and watch a call become a markdown file. both are simulations, and the page says so. it's at dictate.jass.gg.

the site has a changelog page, made from these notes.

the readme is back, and it is short: what it is, how to install it, how to update it, and the two jobs.

in the app, the model card in setup now says the model runs on your mac, so there is no wait for a server. the macOS prompt for the transcripts folder says what the app writes there and nothing else.

nothing else in the app changed. everything else is 0.11.1.

a new logo, and nothing else.

the mic is the head now. it is a bald gold microphone in aviators. the old badge had a head, a mic for a jaw and sound waves on both sides, and in the menu bar all of that shrank to a gold smudge. you can tell this one is a mic in sunglasses at any size.

the menu bar gets a simpler cut of the same drawing, with no rim of its own. a gold rim up there still means a meeting, and now nothing else wears one.

the app icon, about, setup and the site all carry it. everything else is 0.11.0.

the meetings release. i recorded my meetings with another tool because i did not trust this one with them. that is over.

a meeting is never lost. two calls back to back each get their own file. stopping a few seconds after starting keeps the words you said. a quit, or a brew upgrade, waits for the transcript. a crash leaves the audio on disk and the next launch writes it out, with its gaps, instead of pretending it was whole. nothing is ever deleted on a hunch: audio the app cannot read is set aside, not thrown away, and try again in history brings it back.

a transcript is never hollow. before any audio goes, the transcript is held against the speech that was heard. one that comes out thin is read again from the audio, and if it is still thin the file says complete: false and why, and keeps its audio until you delete it. a meeting where nobody spoke says so.

the audio waits a day. after the file is written the recording is kept on your mac, compressed, for a day by default (a setting: at once, a day, a week), then deleted by itself. while it is there, transcribe again with on the row redoes the meeting with another model, and the file is replaced in place.

it hears the whole mac. there is no app to pick: record a meeting records whatever the mac plays plus your mic. a call that moves from zoom to a slack huddle stays in one recording.

it reminds you. when a call app has your mic and is playing audio, a pill offers record, once, for fifteen seconds. when the call ends it asks once whether to stop. nothing starts or stops by itself, ever. an optional hotkey starts and stops a meeting too.

your mic follows you. airpods connecting, leaving or dying mid-call move the meeting to the mic you are on, without a tone, and without a gap while the old mic is still delivering; a mic that dies leaves only the second or two nothing came, marked in the file as a gap where the audio really stopped. a mic that comes up and delivers nothing at all is tried on the built-in mic before the start is called off. a mic that goes silent while the call talks is said by name, and stays said until it is heard again. a mic you muted is not a fault.

silence is not damage. two minutes of a quiet far side used to rebuild the tap, chirp through your speakers and mark the file incomplete. now a quiet far side is asked with a quiet tone the tap can hear, and a tap that really died is rebuilt beside the mic still recording your side, with the gap marked from where the far side was last heard. the lid closed on a meeting is a gap by the audio's own clock, not a file that pretends the hour never happened. a tap that cannot be rebuilt is a problem the meeting records through, your side still being kept, and no window opens over your call.

who said it is certain. your side and theirs are transcribed apart, each stretch of speech once, so you and them are the side that spoke, not a guess by loudness. on the mac's own speakers, their voice coming back through your mic is dropped before it is written down as yours; when it is not clear, the words are kept. a silence of a few seconds starts a new paragraph.

three models. parakeet, the dictation model, with nothing to download, for english and european languages. whisper large, the default, for most languages written in english. whisper turbo, written as spoken. record with picks one for a single meeting. whisper no longer swallows a short "yes" or "okay", and no longer writes "thank you" over a silent mic.

the end is quick. the speaker split works through the meeting a minute at a time, so a file lands seconds after stop whether the meeting ran ten minutes or three hours, and memory no longer grows with the meeting.

you always know the state. the menu bar badge wears gold corner brackets when a call is on and not recorded, a partial rim while getting ready, a full rim while recording, and a red triangle when something is wrong; the first line of the menu says the same in words. the lamp holds a calm light during a meeting that only you can see: the hud and the live transcript are hidden from screen sharing.

every meeting leaves a record, with no words in it: how it ended, speech and words per side, gaps, what the mic and the tap went through, the coverage result. copy diagnostics includes the last twenty.

the transcript folder is for agents too. one paragraph per speaker turn; front matter with ended, speakers, words and a reason when a file is not complete; a README.md beside the month folders that explains the layout to whoever reads it cold.

and the model's small parts come down with it at setup: the tokenizer, the speaker-split models, the voice model (now inside the app). a meeting never waits on the network.

the trust release. two bugs made restarting the app the only fix. both are gone, along with everything shaped like them.

plugging in a monitor, closing the lid or waking the mac no longer leaves fn loading forever. the mic is reopened fresh after any change to your audio, displays or sleep, and nothing to do with audio runs where it can freeze the app.

airpods taken by a call no longer leave a lamp that pastes nothing. each press uses whatever mic your mac is on right now: airpods when they're connected, the built-in one when they're not. if the mic changes mid-sentence, what you said up to then is pasted.

only you throw a dictation away. esc still does. sleep, the lock and the five-minute cap end a dictation and keep it. the screen stays awake while the mic is on, and if it locks anyway, the text is on your clipboard when you come back.

the start sound and the lamp now mean the mic is hearing you, not just that you pressed a key. a mic that sends nothing says so by name, no sound from airpods pro, instead of pretending to listen. a speech model that stops answering gets restarted, and nothing waits forever.

the menu tells you when there's a new version, and clicking update to … runs brew for you. to know about one, the app asks dictate.jass.gg once a day, sending only the version you're running. i count those asks per version per day and keep nothing else. settings › general has the switch. if you're on 0.9.2, this is the last brew upgrade you'll have to type.

fix the same misheard word twice and it's learned. learned: jass.gg says so once, and one click takes it back.

copy diagnostics in the menu copies how your last fifty dictations ended. it has none of your words in it, and it's what to send me when something breaks.

and it's quicker. the speech model wakes up when you press, so it's ready when you let go.

the lamp. the gold line at the bottom of the screen is now a glass tube: a translucent body that takes the window behind it, a pale rim, a shade, a halo that spills when you speak. it waves to your voice and lights up with it, and it holds over a white page as well as a black one — the old line vanished on a document. a locked recording pins it with a bead at each end.

it is drawn by hand, not Liquid Glass, because Liquid Glass draws dimmed in any window that is not the key window, and the lamp's panel must never be key — it would take the keyboard from the app you are dictating into.

the pill that carries an exceptional sentence sits on the same invisible stage as the tube, so the window no longer changes shape around it.

the mic-toggle sound plays 8 db quieter. same switch, further from the ear.

the polish release. a hundred small things a switcher would have tripped on in week one, most of them found by reading the code against the wispr flow bar and then fixed one commit at a time.

the key: granting accessibility makes fn work without a relaunch, setup asks you to press your key before it lets you finish, and esc, the double-tap lock and the five-minute cap are written down under key in settings.

what lands: two dictations in a row get a space between them, a dictation into the middle of a sentence keeps its lowercase, "7 p. M." and "1. 5" are gone, and one stays a word. the settings tile measures to the keystroke now, not to the clipboard restore that ran after it.

failure has a voice: every copied pill ends in ⌘V to paste, pills stay up long enough to read, a stray tap is silent instead of chiming twice and saying heard nothing, airpods arriving mid-sentence says so, and try that again in the menu re-runs a dictation the model dropped.

teach it a word: fix a word… points at what the model actually heard, the dictionary runs before the parsers so an entry always fires, importing asks before it replaces your words, and settings lists the words it keeps mishearing.

meetings: transcripts moved out of ~/Documents (icloud syncs it) into ~/andrew-dictate, files are 0600, a saved meeting gets a notification with show in finder, a mac that slept mid-call writes a gap instead of complete: true, and quitting mid-meeting waits for the file.

updating is one command, and every surface says it: brew upgrade --cask jassuwu/tap/andrew-dictate. your gatekeeper approval and your microphone and accessibility grants come with you. about notices when the new version is already sitting in /Applications, and restarts onto it. a second copy of the app refuses to run.

the readme stops recommending right-click → open, which macOS removed two versions ago, and the site finally shows the thing working.

a release-build fix in the about window's credits. nothing you can feel changed.

the site and the readme carry both jobs, dictation and meetings, in the app's own voice. the pipeline lost a stage it had stopped using.

meetings. pick an app from the menu bar, talk, press stop, and get one markdown file with the speakers split out. your mic is you, the app you picked is them.

whisper large-v3 does the listening, on your mac. the live panel shows it arrive — confirmed in ink, the tail dimmed. no audio is kept once the file is written.

setup asks which job you want now — dictation, meetings, or both — and only asks macOS for what that job needs.