Better with every release
Changelog
What’s new, what’s fixed, and what’s coming next.
v0.1.6Nov 13, 2026
Added
-
Calendar, a new page in the sidebar. Click Connect Calendar to allow access, and it shows the calendars already on this Mac, such as iCloud, Google and Exchange, by Day, Week or Month, like the Calendar app. Week is the default. Day and Week lay out the hours with all-day events on top and a line at the current time, and Month shows up to three events a day. Use the arrows or ⌘← and ⌘→ to move, Today or ⌘T to come back, and click a day to open it. Click an event to see its time, calendar and invitees. A meeting with a video link has Join, and one that starts within 15 minutes or is under way has Take notes, which starts meeting notes with that event's title and invitees. The page says how many calendars it reads and from which accounts, and Choose calendars opens Settings to leave some out. Add Account…, on the page and in Settings > Calendar, opens Internet Accounts in System Settings, where you sign in to Google, Outlook, Exchange, Yahoo or iCloud with your email; the calendars appear once this Mac syncs them. For Outlook, choose Microsoft Exchange. BetterWispr never sees your password. It updates when your calendars change. If access is off, the page says so and opens System Settings; if this Mac has no calendar accounts, it offers Add Account…. BetterWispr only reads calendars and never changes them, and declined events aren't shown.
-
Hold only, a third shortcut behavior in Settings > Shortcut. Dictation runs only while you hold the shortcut and finishes when you let go. A quick tap no longer starts hands-free dictation: it ends the dictation without transcribing or pasting anything and shows "Keep holding" with the shortcut. Hold or tap stays the default and works as before, and saved settings keep their behavior.
-
Capsule position, in Settings > App under Floating recording capsule. The capsule can sit at the bottom center, bottom left or bottom right of the screen. Bottom center is the default. A meeting prompt set to appear at the top still appears at the top center. Right-click the capsule's controls to move it or hide it.
-
Hide the capsule while you're not dictating by turning off Floating recording capsule in Settings > App, or by right-clicking the capsule and choosing Hide capsule. Hiding it from the capsule shows a toast that says where to turn it back on and offers Undo.
-
Dictate and format, a second shortcut in Settings > Shortcut that is fn ⌃ (hold fn and Control together) by default, the key pair other dictation apps use for their second mode. Like fn alone, it needs Accessibility. Record another shortcut there, or click Reset to go back to fn ⌃. If your dictation shortcut is fn alone, press ⌃ before fn. It works like your dictation shortcut and follows the same hold or toggle behavior. It sends the English dictation to the cleanup provider chosen in Style, which must be Claude Code, Codex, an API connection or Ollama (Apple Intelligence and the built-in model aren't used for it; until you choose one, fn ⌃ shows an error that opens Style), shaped for the app you are in:
- email gets a greeting and paragraphs, with a sign-off only when you say one or usually write one
- work chat is short, with code and identifiers in backticks
- personal chat stays relaxed
- AI prompts stay literal and plain
- documents get paragraphs and lists
- other apps get the usual cleanup, with a notice that the app has no layout yet
Start with a request such as "write an email to Sam saying …" or "… with this content …" and the rest is written as that email, without the request words. The request can also follow a few words, as in "okay so please write an email to Sam …", and questions you ask in the email body are kept. In Notes, TextEdit, a browser page or an app with no layout, a spoken email request is laid out as an email too. If the provider's edit can't be used, a simple greeting and body are laid out on this Mac instead. Gmail, Outlook on the web, Proton Mail, Fastmail, Yahoo Mail, Slack, Teams, Google Chat, WhatsApp Web, ChatGPT and Claude in a browser count as their kind of app. No facts are added that you didn't say. If the provider isn't ready or you spoke another language, the dictation gets the usual cleanup and a notice says why. While Save dictation history is on, your edits of a formatted dictation are kept on this Mac, at most 8 per kind of app. A few of them guide later formatting there as style references, not content. Settings shows how many are kept and can forget them. Your normal dictation shortcut is unchanged.
-
Cohere Transcribe, an on-device speech model in Models for English, French, German, Spanish, Italian, Portuguese, Dutch, Polish, Greek, Arabic, Japanese, Chinese, Vietnamese and Korean, with punctuation. It is a 2.2 GB download that runs on the Neural Engine and needs macOS 15 or later, so it isn't listed on macOS 14. It can't detect the spoken language: choose it in Settings > Spoken language, where the automatic choice reads System language and uses the first of your Mac's preferred languages it supports, or English. Vocabulary isn't used as a hint. The first transcription after the model loads can take about a minute while macOS prepares it, which BetterWispr does in the background after loading. Requested in #36.
-
Dictation languages now works with Parakeet. Under Settings > Microphone and language, with Spoken language on Detect automatically, check the languages you speak in the list that opens. When they all use one alphabet, Parakeet never writes letters from another, so English and Portuguese no longer turn into Cyrillic words. Choosing languages with different alphabets, or All languages, keeps every alphabet as before. Parakeet TDT v3 and Parakeet Ultra filter this way; the other Parakeet models ignore the list. Words from another language written in the same alphabet, such as Polish, can still appear. The
--languagesoption of the command-line tool applies to Parakeet too. -
Add interface languages through native translation folders, with a contributor tool to prepare templates, check resources and package the declared languages.
-
French interface, app status messages and permission descriptions follow the macOS app language, with English as the fallback. Native plural rules and localized keyboard labels preserve model identifiers, saved data, prompts and user-provided names.
-
Summary language for meeting notes, under Models > Meeting notes and in the Notetaker settings menu. Same as meeting is the default: notes are written in the Spoken language set when the meeting was recorded, or, with automatic detection, in the language spoken most in the transcript. Before, notes were usually written in English. Choose a language to always write notes in it, with every notes provider. If Apple Intelligence can't write in that language on this Mac, Models and the meeting's Generate summary card say so, and no summary is written until you pick another language or notes model. Ask answers are unchanged.
-
Write numbers as digits, in Settings > After you speak and on by default. Every spoken number becomes digits, in the pasted text and in Parakeet's live preview, with any speech model, in English and in every other language written with spaces between words: "two cats" is typed as "2 cats", "quatro mil quinhentos e sessenta e sete" as "4567" and "doscientos cincuenta mil" as "250000". Numbers spoken one after another join when there are three or more, when they sound like a year, or after "number", "PR", "issue" or "ticket", so "PR número quatro cinco seis sete" becomes "PR número 4567", "call five five five" "call 555" and "nineteen ninety nine" "1999"; a later number of 100 or more starts a new one. Two other numbers in a row stay as words, so a time such as "eleven thirty" or "son las dos treinta" and "fifty fifty" are kept as spoken. Outside English no thousands separator is added. A decimal or grouping mark is written only where it is said between numbers, so "três vírgula cinco" becomes "3,5" and "zero ponto um ponto seis" "0.1.6". "One", "um", "une" or "ein" on its own stays a word, since it is usually an article or pronoun, unless it follows "number", "PR", "issue" or "ticket", and "thousand" or "mil" on its own stays a word too. Number words for languages other than English come from macOS, and the language comes from Settings or, on Auto, from the text. In English this replaces keeping small numbers as words, so "one or two" is typed as "one or 2"; English amounts keep their separators, such as "250,000", except after "number", "PR", "issue" or "ticket". Turn it off to keep every number as spoken words. Medium cleanup may still write numbers as digits.
-
Mute your microphone during a meeting note. Click the microphone button in the meeting toolbar to mute it, and click again to unmute; the arrow next to it still chooses the microphone. While muted, the button turns red and only other apps' audio (Them) is recorded and transcribed. What you said just before muting is still transcribed. Each new meeting starts unmuted.
-
Send keys, in Settings > Send keys and empty by default. Add an app and record the key that submits in it, and BetterWispr presses that key after pasting a dictation there, so dictating into a chat or agent client that Send to Claude cannot recognize still submits hands-free. A bare Return, Tab or ⌘Return all work. The recorder refuses ⌘Q, ⌘W and Delete or Forward Delete, with or without other modifiers, saying why, and refuses your dictation shortcut, which would clash with the send. A modifier on its own is refused because it submits nothing, and Esc leaves the recorded key as it is rather than being recorded. Terminal and iTerm2 aren't offered, because a key there could run a shell command; Send to Claude still covers Claude Code in them. Each app has its own switch, so a rule can be turned off without losing its key, and a rule for an app that is no longer installed says so instead of disappearing. The capsule shows ↵ for Return, or a keyboard glyph for anything else, and its tooltip names the key. Unlike Send to Claude, a send key fires wherever the cursor is in the app rather than only in a field BetterWispr recognizes, so pick a key that cannot do damage in the app's other fields; hold Shift as you dictate to paste only. The wait before sending, Don't send, "Pasted, not sent" and taking over a pending send by dictating again all work the same as for Send to Claude. Needs Accessibility. Contributed by Mike Lyons (frenchie4111).
-
Allow any Claude Code sign-in for meeting notes, under Models > Meeting notes when Claude Code is the notes provider, off by default. Claude Code normally needs a Claude subscription sign-in. With this on, meeting notes also work when it is set up with an API key, a cloud provider, or a gateway or proxy. Models shows where Claude Code says it will send meeting text, including a custom base URL in its settings files, and nothing is sent until you confirm that destination. If the destination changes, BetterWispr asks again before sending. Each request is capped at an estimated $1, and the summary names the model and destination that wrote it. Ask, Medium cleanup, screen and clipboard context and correction review don't use this switch: they still need a Claude subscription, and Medium keeps Light cleanup otherwise. Without it, a Claude Code that is signed in another way says so and points to this switch.
-
Mute other apps while dictating, in Settings > App and on by default on macOS 14.2 and later. It silences browsers, Spotify and every other app while the microphone is recording, so a video in Safari or Chrome or a song in Spotify goes quiet while you speak and is heard again when you stop. Other apps are muted rather than paused, so videos keep playing silently. BetterWispr’s own sounds still play, and your volume and mute settings are not changed. Nothing is muted while meeting notes record, or when a browser or a known call app such as Zoom, Teams or FaceTime is using the microphone as the recording starts. A call in any other app goes silent while you speak. If muting fails, a message says so and the dictation carries on. Whether other apps come back on their own if BetterWispr crashes during a dictation hasn't been tested. The first dictation that mutes asks for System Audio Recording, and nothing goes silent until you allow it. If you turned the setting off, turning it back on asks right away. A workspace saved before this change has Pause music while dictating and Mute other apps while dictating turned on once, since both used to be off by default; turn either off in Settings > App and it stays off. Pause music while dictating is a separate setting and works as before.
-
English Medium cleanup infers bullet lists and numbered steps from the meaning of the transcript, without requiring “list”, “shopping list”, ordinals or bullet commands. Natural repetitions and mixed transitions can form a list of three or more items, or two when you number them; connected narratives and comparisons remain prose, and "new line" and "new paragraph" still break the text. Uses the selected cleanup model, including Apple's on-device model, with a second call only when a dictation looks like a list. Selected-text cleanup never forms a list. Light cleanup keeps its filler and stutter removal without guessing list structure.
-
Quitting BetterWispr in the middle of a dictation keeps it. A dictation that is still recording or transcribing is saved to Failed Dictations, and History lists it under Not transcribed after the next launch. Before, quitting deleted it.
-
Keep a dictation that couldn't be transcribed. When the speech model or an API connection fails, BetterWispr keeps the recording instead of deleting it, and Try again on the error card transcribes that same recording with the speech model selected now, rather than starting a new dictation. The audio is also copied to
~/Library/Application Support/BetterWispr/Failed Dictations, readable only by your macOS user, so it survives closing the card and quitting. History lists these under Not transcribed, where Transcribe opens one on the Transcribe page, the folder button shows it in Finder and the trash button deletes it. A recording is deleted once it is transcribed, after 7 days or a shorter Keep history age limit, or when more than 20 are kept or they pass 1 GB in total. The oldest go first, and the recording that just failed is always kept, even if it is larger than 1 GB on its own. The folder is left out of backups. Cancelled dictations and ones with no speech are not kept, nothing is copied to disk while Save dictation history is off, and turning it off deletes what is kept. When the selected speech model is an API connection or Codex, the button reads Send to and names where the audio goes, and a failing local model never falls back to one. -
Keep history, in Settings > App under Save dictation history. Choose Forever, the latest 100 dictations, 3 days, 2 weeks or 3 months. Older dictations are deleted from this Mac when BetterWispr opens and after each dictation, and a toast at launch says how many went. A choice that would delete dictations asks first, says how many and warns that usage totals and Insights will shrink, since they count saved dictations. It is Forever by default, so nothing is deleted until you choose. While Save dictation history is off, Keep history deletes nothing. If the Mac's date moves ahead by more than a day since BetterWispr last checked it, age limits wait until the new date has held for a day, so a clock set ahead by mistake doesn't delete dictations. A clock left wrong for longer than that is treated as correct.
-
Add a space after pasting, in Settings > After you speak and off by default. Each pasted dictation ends with a space, so the next one doesn't run into it. It is skipped when the text ends in Chinese or Japanese, or already ends in a space or line break, and History keeps the text without it. A dictation that joins one waiting to be sent to Claude uses that space instead of adding its own.
-
Sound volume, in Settings > App under Dictation sounds. A slider sets how loud the start, finish, cancel and needs-attention sounds play, and plays the start sound when you let go. It starts at the level the sounds always used.
-
Notify when a word is learned, in Settings under Learn from my corrections and on by default. Turn it off to stop the "Learned …" toast after you fix a dictation. Corrections are still added to Vocabulary, where entries marked Learned can be reviewed and removed. Without the toast there is no Undo button.
-
Send to Claude, in Settings > After you speak and off by default. After pasting, BetterWispr presses Return when the cursor is in Claude Desktop's message box, the Claude Code panel in VS Code, or a Terminal or iTerm2 tab where Claude Code is running, so a dictated prompt is submitted hands-free. Anywhere else, such as a shell or a file, the text is only pasted. A tab counts only while a Claude Code process is in front in it: Terminal and iTerm2 are asked which tab is focused once you allow it. VS Code terminals and Ghostty paste without sending because a stale title cannot prove which process receives Return. Return also answers a question Claude Code is asking, such as a permission prompt. The capsule shows ↵ when a dictation will be sent. Hold Shift as you finish, or start with Shift and your shortcut, to paste only. After pasting, the capsule counts down for a second: dictate again to add more, and the next dictation is pasted after the first and sends both, or press Escape or click Don't send. Wait before sending changes the countdown, from 120 ms to two seconds. Return is skipped, with "Pasted, not sent", when you switched apps or fields. In VS Code this turns on its accessibility support, which may ask once whether to optimize the editor for a screen reader; No is fine.
-
Tell me when no speech is heard, in Settings > App and on by default. Turn it off, and a dictation that ends without words closes quietly instead of showing "No speech heard." with a sound. The card still appears when an earlier dictation was waiting for this one to send it, since that dictation stays pasted but not sent.
-
Codex · Subscription under Models > Subscription transcription. Explicit selection sends audio directly to an unofficial ChatGPT endpoint using your local Codex sign-in. No separate transcription server or helper is required. BetterWispr does not store or refresh credentials. Local defaults and notes providers stay independent. See setup. The developer CLI accepts
--model codexfor batch transcription. -
Transcribe audio and video files and web links from the new Transcribe page in the dashboard. Drop or choose an MP3, M4A, AAC, WAV, AIFF, CAF, FLAC, MP4, MOV or M4V file, or paste a link such as a YouTube video. BetterWispr transcribes it on this Mac with the model you pick on the page, or with an API connection only if you pick one. Links download with yt-dlp, a free open-source downloader. Nothing downloads until you click Install yt-dlp, and BetterWispr checks it against its published checksum first. YouTube links work best with Deno installed (
brew install deno). Long recordings are cut at the quietest moment every 30 to 60 seconds. On three TED videos, cutting this way got 1,233 words wrong out of 12,592, against 1,282 when sending the whole file at once and 1,335 when cutting it like meeting audio. Transcripts keep their timestamps, can be copied or exported as plain text or Markdown, and label Speaker 1, Speaker 2 when Speaker detection is installed.BetterWisprCLI --media PATH|URLdoes the same from the command line, and--download-yt-dlpinstalls yt-dlp. -
A built-in notes model that runs on this Mac. Download Qwen3.5 2B (1.28 GB) under Meeting notes in Models, and it writes meeting notes, edits Medium dictation and answers Ask without Apple Intelligence, Ollama, a CLI or an API key. It supports macOS 14. It uses about 1.5 GB of memory while it works, most of it the model file read from disk, and frees it after five idle minutes. BetterWispr uses it only when you choose it, and it never downloads on its own. A finished download makes it your notes provider unless you picked another one meanwhile. A 2B model can repeat points or give a task to the wrong person, so check its notes.
BetterWisprCLI --download-builtin-modelinstalls it, and--notes-model builtinuses it with--meetingand--bench. -
Ask questions about one meeting from its Ask tab, or about all your meetings from the Notetaker page. Answers cite the transcript lines they came from, and clicking a citation opens the meeting at that line. Ask sends your question and the meeting lines that best match it to the notes provider you pick for Ask, or to the one that writes your summaries when you keep Same as notes. A meeting that is still being processed can be asked about from its transcript so far.
-
Ask keeps your questions and answers. A meeting's Ask tab shows the ones about that meeting, and the Notetaker page shows the ones about all your meetings, after you quit and reopen BetterWispr. Each answer names the model that wrote it. Clear removes them from this Mac after you confirm. Deleting a meeting also removes the answers about all meetings that cite it. A question that fails or is cancelled isn't kept.
-
What did I miss? in the Ask tab of a meeting that is recording. It answers from the last five minutes of the transcript, with citations, and leaves out My thoughts.
-
Filters for Ask on the Notetaker page. Ask about meetings from today, the past 7 days or the past 30 days, and about meetings with one Calendar invitee.
-
Provider presets for a new notes API connection in Models. Pick OpenAI, Anthropic, Google Gemini or OpenRouter to fill in its address and a fast model, then add your API key. Anthropic says its OpenAI-compatible API is not a long-term or production-ready solution.
-
Calendar for meeting notes, in Settings > Calendar and off by default. Turning it on asks macOS for Calendar access. A note started during an event, or up to five minutes before it, takes the event's title and lists the people invited under its date. BetterWispr reads the calendars already in the Calendar app on this Mac, from every account including shared and subscribed calendars, so it works offline and needs no sign-in. Each calendar has its own switch, and calendars added later are included. Declined and cancelled events are skipped. When events overlap, the one with a Zoom, Google Meet, Microsoft Teams or Webex link wins. The event title and invitees go to your notes model with the transcript. Invitee names also help the speech model recognize them.
-
Connected apps, in Settings and off by default. Turn on Share meetings with AI apps, and Claude Desktop, Claude Code, Cursor and other apps that support MCP can list and search your meetings and read their summaries, action items, your notes and transcripts with speaker names and timestamps. Settings shows the line to run for Claude Code, the snippet to paste for Claude Desktop and Cursor, and the command for other apps, each with a Copy button. Connected apps can only read, never change or delete anything. BetterWispr checks the setting on every request, so turning it off stops sharing right away without restarting the other app. BetterWispr itself sends nothing over the network, but an app you connect sends what it reads to its AI provider. Long meetings arrive in parts, and files from the Transcribe page are not shared.
-
Correct the speakers after a meeting. Click the speaker name on any line to give it to another speaker or a new one, or to move one of your lines back to Me. Use the menu beside a speaker's name field to merge that speaker into another. A merged speaker keeps its target's name, or takes the other name when the target has none. A merge can't be undone, but you can move lines back one at a time. The speaker menus appear once the meeting has finished processing.
-
A meeting recorded without Speaker detection offers to download it. The download applies from the next meeting, and you can dismiss the offer for good.
-
Setup offers Speaker detection below the speech model. It is optional, and nothing downloads until you click Download. If you click it while the speech model is still downloading, Speaker detection starts once the speech model is ready.
-
Name a speaker after someone invited to the meeting. When Calendar added the meeting's invitees, the button beside each speaker's name field lists the invitees no speaker has been named after yet. Pick one to fill in the name.
-
Tell apart people in the room, in Models > Meeting speakers and off by default. With Speaker detection installed, it also separates the voices on your microphone, so people sitting with you get their own speaker numbers next to the people on the call. The voice heard longest on your microphone stays Me, not counting the call's sound picked up from your speakers. If the wrong voice ends up as Me, click the speaker name on a line to move it. Your microphone audio stays on this Mac until the speakers are found, then BetterWispr deletes it. How well it separates people in a room has not been measured.
-
Speaker labels during a call. With Speaker detection installed, the lines from the other side of a call show Speaker 1, Speaker 2 and so on while the meeting records, a moment after each line appears, where before they stayed Them until the meeting ended. BetterWispr listens again to the last five minutes of call audio on this Mac after each new line, using the Speaker detection models already downloaded, so nothing new downloads. The labels are a first guess: a line's speaker can change while the meeting records, one person can briefly show as two speakers, and someone who was silent for more than five minutes can come back under a new number. The pass after the meeting replaces them and keeps each number with the voice that mostly had it. Your microphone lines stay Me during the call, and without Speaker detection nothing changes. Measured only on synthetic conversations spoken by macOS
sayvoices, not on real calls: over seven two-minute conversations with one to four voices, 48% to 80% of line time had the right speaker when first shown and 92% to 100% when the call ended, and three of the seven showed one speaker more than there were. Over two ten-minute conversations with three voices, one had 65% right when first shown and 94% at the end, and the other 54% and 63%, where the pass after the meeting also merged two of the voices. Each listen took about 0.9 seconds on an Apple M4 Pro with five minutes of audio. How much it slows transcription during a real call has not been measured. -
My voice, in Models > Meeting speakers. Click Record My Voice and talk for 30 seconds, and BetterWispr saves a profile of your voice and deletes the recording. When Tell apart people in the room is on, the microphone voice closest to your profile stays Me, even if someone beside you talks more. Without a profile, or when no voice in the meeting is close enough to it, the voice heard longest stays Me as before. A sample with less than 20 seconds of speech is refused, because a shorter one gave a profile that matched nothing. The profile is stored with remembered voices in
Voices.json, readable only by your macOS user, is never uploaded, and Time Machine backs it up. Remove deletes it, and Forget Voices deletes it too. Measured only with macOSsayvoices: with one of three voices on the microphone recorded as My voice, 92% to 100% of microphone time was rightly marked Me or not across six choices of voice. A synthetic voice sounds the same in every recording, so real voices will match less often, and that has not been measured. A remembered voice can't be marked as Me yet. -
Remember named voices, in Models > Meeting speakers and off by default. When it is on and you name a speaker, BetterWispr saves a profile of that voice with the name. In later meetings, when a speaker's voice is close enough to a saved profile, the speaker's name menu offers "Sounds like" with that name as its first choice. No name is filled in until you pick one, so the transcript and the notes show only names you chose. If you give that speaker a different name and then press Return, click away from the name field or pick an invitee from the menu, BetterWispr deletes the profile that made the suggestion. A name you type and erase deletes nothing. Close enough means within the cutoff SpeakerKit uses to group one person's speech inside a single meeting. That cutoff was not tuned for matching across meetings, and how often the suggestion is the right person has not been measured. A voice profile can identify a person, and privacy laws such as the GDPR and Illinois' BIPA can treat it as biometric data, so ask people before you name them. Profiles stay in
~/Library/Application Support/BetterWispr/Voices.json, readable only by your macOS user. They are never uploaded or written into meeting files, but Time Machine backs them up. Forget Voices deletes them all. Deleting a meeting deletes the profiles named in it, and the delete confirmation says so. Name speakers before you quit BetterWispr. A meeting's voices are kept only in memory, so a name given after a relaunch isn't remembered. The profiles come from the Speaker detection models. Their CC BY 4.0 license sets conditions only on sharing the models, and BetterWispr shares neither the models nor the profiles. -
Press Escape to cancel a dictation from any app while it is starting, recording or transcribing. Escape also works with your shortcut's Option, Control or Shift still held, but never with Command, so Command-Option-Escape stays Force Quit. Undo still appears when audio was captured. Escape goes back to the app you're in as soon as the dictation ends. Hovering the capsule while you dictate shows "Esc to cancel".
-
An optional second shortcut for cancelling a dictation, in Settings > Shortcut. It is off by default and Escape keeps working. Record a combination with a key, such as ⌃⌥K or F5, and it cancels from any app while a dictation is starting, recording or transcribing, without pasting it. A modifier key alone can't be used, and neither can your dictation shortcut. The combination works as usual in other apps whenever you aren't dictating. Clear turns it off.
-
Clean up selected text with an optional shortcut, in Settings > Shortcut. It is off by default. Select text in any app and press the shortcut, and BetterWispr edits it like a Medium dictation with the cleanup provider from the Style page, then pastes the edit over your selection. The selected text goes to that provider, as a Medium dictation does. It needs Accessibility and at least four words of English, and it never reads a password field. If you switch apps or change the selection before the edit is ready, nothing is pasted and the edit goes to the clipboard. A toast says when the provider changes nothing or its edit is refused, and Cancel on the progress toast stops it. Starting a dictation stops it too, so the two never paste at once. A modifier key alone can't be used, and neither can Escape or your dictation or cancel shortcut. Clear turns it off.
-
Fn (🌐) can be the dictation shortcut on its own. In Settings, click the keyboard shortcut and press and release Fn. Like other single modifier keys it needs Accessibility, and it only starts dictation when no other modifier is held. Set "Press 🌐 key to" to "Do Nothing" in System Settings > Keyboard so macOS doesn't also show emoji or switch the input source.
-
Choose a cleanup provider for Medium cleanup on the Style page. It is Same as notes by default, so Medium keeps using your notes model until you pick another, such as a small Ollama model for dictation while a larger one writes meeting notes. Apple Intelligence, the built-in model, Ollama, Claude Code, Codex and API connections can be picked, and an API connection is used only when you pick it. If the provider you pick isn't ready, Medium dictations get Light cleanup.
-
Pause music while dictating. A new setting pauses Music and Spotify while the microphone is open for a dictation and resumes only what it paused when you stop or cancel. It is on by default and asks macOS for permission to control those apps when BetterWispr starts or when you turn it on, for each one that is open and hasn't been asked yet. Together with Mute other apps while dictating, Spotify and Music pause, and a video in your browser goes silent, while you speak.
-
A start sound plays when the capsule begins listening, so you know the microphone is open before you speak. It is the macOS Tink sound, and a
started.wavfile in~/Library/Application Support/BetterWispr/SoundOverridesreplaces it. It stays quiet when a dictation fails to start, while a meeting is being recorded, and with Dictation sounds off. -
Choose each dictation sound in Settings > App. With Dictation sounds on, Start, Finish, Cancel and Needs attention each list the macOS system sounds, and picking one plays it. A sound you don't pick stays Tink, Purr, Bottle or Basso. A file in
SoundOverridesstill wins over the picker, and that picker says so and is greyed out. Previews stay quiet while a meeting is being recorded. -
Hide capsule while dictating, in Settings > App and off by default. The capsule stays out of sight while you speak and while your words are transcribed. Errors, Undo and the copied and finished cards still appear, and the capsule comes back for the last 30 seconds before a long dictation stops. Escape and your cancel shortcut still cancel, and the menu bar icon turns into a microphone while you speak. Show words as you speak is off while the capsule is hidden.
-
A Cream theme, in Settings > Appearance, for a warmer page than Light's pure white. It paints the dashboard, the welcome guide and the Notetaker panel cream with slightly lighter cards, and keeps the system text colors and controls. Body text passes WCAG AA contrast. Grey secondary text is about 3.9:1, the same as in Light.
-
Unload speech model when idle, in Settings > Microphone and language. Choose After 5 minutes, After 15 minutes or After 1 hour to free the speech model's memory between dictations. The next dictation loads the model again from disk, so it takes a moment longer to finish. It is Never by default, the timer never runs out during a dictation, and waking or unlocking the Mac doesn't load a freed model back. The built-in notes model keeps its own five-minute unload.
-
Each speech model in Models lists the languages it supports by name, in your Mac's language. A model with one language shows its name, and a model with more shows a count, such as "25 languages", that opens to the full list.
-
Dictation languages, in Settings > Microphone and language, shows with a Whisper model and Spoken language set to Detect automatically. Pick the languages you speak, such as English and Spanish, and detection chooses the most likely of them instead of any language Whisper knows. With languages picked, the first 30 seconds of a dictation decide its language, which takes one more pass over that audio before transcribing. All languages, the default, detects as before. Other models ignore the list.
BetterWisprCLI --transcribe-file PATH --languages en,esdoes the same. -
Chinese script, in Settings > Microphone and language. Choose Simplified or Traditional to get Chinese dictation in those characters, whichever script the speech model wrote. It shows when Spoken language is Chinese, or Detect automatically with a model that supports Chinese, where it converts only dictations recognized as Chinese, so Japanese stays as it is. The conversion uses the converter built into macOS and runs on this Mac. As recognized, the default, leaves the characters alone, and History's Original transcription keeps what the model wrote.
-
The dashboard sidebar shows the app version with a What's new link to the changelog. About has new Changelog and Support rows, and Support opens Buy Me a Coffee.
Changed
- Home's greeting card is a plain card now, without the colored glow in its corner.
- Settings fills the width of the window instead of staying in a narrow column in the middle.
- Setup recommends Parakeet Ultra and lists Apple on-device speech only when it supports your language on this Mac. If Apple is your selected model and it isn't available, setup says so, keeps your setting and holds Continue until you pick another model. Models marks Apple as unavailable and turns off its Use button.
- Dictation no longer stops at two minutes. It used to stop at two minutes without warning. Now it records until you release or press your shortcut, with a brief reminder every hour, up to a limit of 4,278,190,080 audio frames: about 24 hours 45 minutes at 48 kHz, half that at 96 kHz. At the limit it finishes by itself and transcribes what it captured. The live preview still stops following after ten minutes, and the full recording is transcribed when you finish.
- A recording of ten minutes or longer now asks before it is sent to an API connection and before it is pasted automatically. If you don't send it, the audio stays on your Mac and you can try it again. If you don't paste it, the text is left on your clipboard. Local models never ask before transcribing.
- API connections now load only the pieces of a long recording that are being uploaded, instead of holding the whole recording in memory first.
- A long dictation sent to an API connection is uploaded in pieces at the same time, up to four at once, instead of one piece after another. The transcript keeps the pieces in order. If the provider answers that it is rate limiting (HTTP 429) or a piece times out, BetterWispr keeps the answers it has and sends the remaining pieces one at a time. A piece that fails then fails the dictation: nothing is pasted, and the recording is kept so you can try again.
- A dictation interrupted by a microphone or recording failure now keeps what was captured, so you can try it again and find it in History when history is on. Before, the audio was deleted.
- Recordings left in the temporary folder by a crash are deleted at the next launch, once they are more than a day old.
- Hold to talk in Settings > Shortcut is now named Hold or tap, and its caption says how to finish a hands-free dictation. It works as before. Hold the shortcut to speak and release to finish, or tap it once to dictate hands-free and press it again to finish.
- With a Whisper model selected, Spoken language in Settings offers every language that model supports, about 100, instead of the shorter list shared with the other models.
- Speaker detection now works on macOS 14. It still needs the call audio the notetaker records on macOS 14.2 or later. It uses new models from Argmax SpeakerKit, about 15 MB, in place of the ones 0.1.5 downloaded. If you installed Speaker detection in 0.1.5, Models shows it as not downloaded, so download it again under Meeting speakers. Once the new download finishes, BetterWispr removes the 0.1.5 files, about 21 MB. Meetings you already recorded keep their speaker labels and names. The speaker accuracy figures in 0.1.5 were measured with the old models and have not been measured again with the new ones.
- Use screen context now also fixes misspelled names in a dictation with the names in the window you dictate into, such as a teammate's name in a chat. It works at every cleanup level, so its switches now show without Medium. It needs Parakeet with the phrase booster and English or automatic language. Text read for spelling stays on this Mac. Only Medium cleanup sends screen context to your notes model, as before.
- Start the next meeting while the last one is still processing. Stopping a meeting no longer locks the notetaker while it transcribes its last chunks, finds speakers and writes notes. The capsule, the menu bar and the dashboard can start a new meeting right away, and the stopped one finishes on its own. Its last words stay in its own transcript, and the dashboard does not jump to its note while the new meeting records. Skip, Cancel and Update summary act on one meeting. While two meetings overlap, each holds its own copy of the speech model.
Fixed
- The floating capsule no longer blocks clicks in apps behind it. Its invisible window covered a large area at the bottom of the screen, so buttons there, for example in apps tiled in split view, could not be clicked. Only the capsule itself and a small margin around it now take the pointer, and hovering it still opens its controls.
- The capsule no longer uses Liquid Glass on macOS 26 and later, where it could turn gray over bright windows and make its white text and icons hard to read. It now uses the same background as earlier macOS versions.
- A dictation started while AirPods are playing for your iPhone or another device no longer records only silence. macOS could hand the AirPods to your Mac a moment after BetterWispr opened their microphone and then never send their sound. When a microphone sends nothing but exact silence for 2 seconds, BetterWispr restarts it once and hears what you say after that.
- Audio interfaces with more than two inputs, such as Focusrite Scarlett interfaces, no longer fail with "No microphone input is available". BetterWispr records the interface's first two inputs, where its microphone inputs usually are, and leaves out loopback channels.
- Stopping a dictation while BetterWispr is busy no longer throws the recording away with "The microphone stopped delivering audio". BetterWispr waited for the last moment of audio on the same thread that draws its windows, so a busy app could miss it even though the audio had arrived.
- Leaving History no longer freezes BetterWispr for minutes when you have many dictations. History built every saved dictation at once, so a long history took seconds to open. Switching to another page then made macOS take thousands of views apart one at a time. History now builds only the dictations on screen.
- Pasting into Screen Sharing works. BetterWispr now presses Command as a key of its own around V, the way a keyboard does. Screen Sharing passes Command to the other computer only when it sees that key, so it used to type the letter "v" there. Screen Sharing also misses clipboard changes made while its window is in front, so BetterWispr first chooses Edit > Send Clipboard, and your text stays on this Mac's clipboard. With Use Shared Clipboard on, Screen Sharing dims Send Clipboard, so BetterWispr turns that setting off, sends the clipboard and turns it back on. This adds about a second before the paste. When Send Clipboard can't be chosen, BetterWispr copies the text and doesn't paste, because the other computer would paste whatever it last copied.
- Automatic paste checks that the field focused when dictation started still has focus. If you move to another field or tab in the same app, BetterWispr copies the transcript for manual paste. Apps that do not expose their focused field through Accessibility keep the existing app-level check.
- Back-to-back dictations no longer restore an older clipboard snapshot over something you copied between them. Clipboard restoration also preserves rich-text formats and an empty clipboard.
- An Ask answer that finishes after you delete a meeting it cites is discarded, so it cannot restore references to that meeting in your saved questions and answers.
- Long meetings no longer fail with "The model couldn't condense this meeting enough". BetterWispr shrinks a long transcript in up to 8 rounds instead of 3, asks for fewer words when the notes model writes more than it was asked for, and when the model still can't shrink the meeting enough, writes notes for each remaining part and joins them in order. Such notes can be long and repeat a point, but every part of the meeting goes into them. The Summary tab also lets you pick another notes model next to Generate summary, so you can retry without going to Models.
- Dictation sounds now play on a new install. With the setting on, BetterWispr stayed silent unless you had added your own sound files, because none ship with the app. Finishing a dictation now plays the macOS Purr sound, cancelling plays Bottle, and a problem that needs your attention plays Basso. A
finished.wav,cancelled.wavorattention.wavfile in~/Library/Application Support/BetterWispr/SoundOverridesstill replaces the matching sound. - Claude Code and Codex are now found when installed through nix or anywhere your login shell can run them, so Meeting notes and Medium cleanup no longer report them as missing.
- Tagalog dictation with Smallest AI no longer fails as an unsupported language. BetterWispr renamed the language's code on the way out, and Smallest AI accepts only the original. Javanese and Norwegian are now sent the same way, with the codes Whisper uses.
v0.1.5Oct 12, 2026
Added
- Meeting notes can tell the people on the other side of a call apart. Download Speaker detection (about 21 MB) under Meeting speakers in Models. When a meeting ends, BetterWispr listens back to the call audio on this Mac and labels each line Speaker 1, Speaker 2 and so on, in the order they first spoke. The notes model gets the same labels, so the summary can say who said what. While it works, the notetaker shows "Telling the speakers on the call apart" with a Skip button that goes straight to the notes. It needs macOS 15 or later and the call audio the notetaker records on macOS 14.2 or later. When the voice changes partway through a line, BetterWispr cuts the line there and transcribes each part again, so a quick reply goes to the person who gave it. On six recorded two-person calls, 93% of words went to the right person, against 85% without the cut, and on four recorded four-person meetings 82% against 77%. The shorter parts cost a little accuracy, about 4 more wrong words in every 1,000, and with an API speech connection each part is sent again. Turns shorter than a second stay with the line around them, and when two people talk at once the part goes to whoever spoke longer. Your microphone is still Me, so people in the room with you are not separated. Meetings recorded before the download keep Them.
- Name the speakers. Fields above the transcript let you type a name for each detected speaker, and the name replaces "Speaker 2" in the transcript, in Copy and Copy transcript, and in notes you generate after that. Renaming marks the summary as out of date so you can regenerate it with the names.
BetterWisprCLI --download-speaker-detectioninstalls the speaker detection models.--meeting PATHplays an audio file through the notetaker's chunking, transcribes it as a meeting does and prints each line with its speaker as JSON, and--meeting LINES.jsonl --notes OUT.mdwrites notes from an earlier run without transcribing again.scripts/fetch_bench_corpus.py --meetingsdownloads four recorded AMI meetings and six EdAcc video calls with who spoke when, andscripts/evaluate_meetings.pyscores the words and the speaker labels.- Writing modes, under Style > Modes. A mode sets its own cleanup, tone and screen context, and with Medium cleanup it can add instructions for your notes model, such as "Use British spelling." Anything left as "Same as default" follows your defaults. A mode applies in the apps you add, on the websites you add, or when you start or end a dictation with its trigger word, which BetterWispr removes from the text. With the trigger "casual", "Let's keep it casual." types "Let's keep it.", so pick a phrase you wouldn't otherwise say, such as "casual mode". A trigger word wins over a website, and a website wins over an app. A trigger word can turn screen context off but not on.
- Website modes match a site and its subdomains, plus an optional path.
github.comalso covers gist.github.com but not notgithub.com, and*.github.comandgithub.com/apple/*work too. They work in Safari, Chrome, Arc, Dia, Edge, Brave, Opera, Vivaldi, Comet, Orion and Yandex Browser, not Firefox. BetterWispr reads the address through Accessibility when you start dictating, so there is no Automation prompt, and it never saves or sends it. While an enabled mode lists a website, the capsule can take up to half a second longer to show recording in a browser, and no audio is lost. - Use screen context, under Style, off by default. With Medium cleanup it sends your selected text, your clipboard and the text of the window you dictate into to your notes model, so it can spell what you see. Your notes model may be a cloud service. Nothing is read while a password field is active, and clipboard items marked as passwords are skipped. Read window text from the screen, also off by default, reads the window from a screenshot when an app shares under 200 characters through Accessibility. It needs Screen Recording, and dictation never asks for it.
- Show words as you speak, in Settings > App and on by default. With a Parakeet model the capsule shows your words while you record, confirmed words bright and the newest guess dimmed, two lines at a time. The pasted text is still the recognition of the whole recording, and the preview never reaches History or correction learning.
BetterWisprCLI --live-file PATH [--model ID] [--language en]plays an audio file through the live preview in real time with a Parakeet model, prints each confirmed and provisional update with the time its pass took, then prints the full transcription.- Spoken language in Settings offers Turkish, for Apple on-device, Whisper and OpenAI-compatible connections. Parakeet models do not list it, and Sarvam and Smallest AI do not support it.
- Show in Dock, in Settings > App and off by default. BetterWispr runs from the menu bar without a Dock icon: your dictation shortcut still works, and the menu bar icon opens the dashboard or quits. Turn it on to bring the Dock icon back. The change applies right away and stays after a restart.
Changed
- Parakeet Ultra is now the recommended model in setup and is listed above Parakeet TDT v3 in Models. Before cleanup, on 300 recorded video-call segments in Indian, Nigerian and US English, it got 14.45% of words wrong against 15.96% for Parakeet TDT v3 (14.95% against 15.66% after Light cleanup), at the same speed, with the same 25 languages. On 597 more Indian English segments it got 11.5% wrong against 12.3% before cleanup, but after Light cleanup the two were level at 12.05% and 12.11%, because cleanup removes more of the fillers v3 writes. On four recorded meetings cut into chunks as in 0.1.4, it got 22.8% wrong against 25.7%. With this version's cut at voice changes Ultra got 23.2%, and v3 was not run that way. Dictation and meeting notes both use it once selected. The model you use now does not change on its own. If you use Parakeet TDT v3 or v2, Home suggests trying Parakeet Ultra and Models marks it Recommended; download it in Models and choose it to switch.
- Learn from my corrections asks your notes model first. With Medium cleanup, English text and a ready notes model, each fix the rules would learn goes to the model with up to three words on each side, and only the fixes it confirms are learned. The model is asked to turn down everyday word swaps such as "meet" to "meat", while a misheard name such as "post hog" to "PostHog" stays allowed. If the model gives an invalid reply or isn't running, BetterWispr learns by its rules as before. The "Learned" toast has an Undo button that removes the words it just added.
- Medium cleanup uses a new edit prompt. Vocabulary is given as the spelling to use, and a mode's instructions go with it. With llama3.1:8b through Ollama, 18 of 18 test edits came back as plain edits, in 0.55 to 1.86 seconds. With Apple Intelligence, lists come back numbered and "no wait" and "sorry I mean" are resolved, but each dictation takes 0.5 to 0.8 seconds longer, and "actually scratch that" stayed in the text in 2 of 2 runs, where the old prompt removed it.
- When a Medium edit times out or hits a passing network error, BetterWispr tries up to three times, allowing 7 seconds plus 1 for every 10 words each time. Notes CLIs get one try of 10 seconds plus 1 for every 10 words. If every try fails, the dictation keeps Light cleanup.
- With Ollama, BetterWispr loads the notes model 450 ms into a recording, at the size the edit needs. When the model wasn't loaded yet, the wait before an edit fell from 0.91 to 0.06 seconds with llama3.1:8b. Apple Intelligence is prepared the same way. Dictation edits ask known API services for low reasoning effort, Claude Code gets
--effort lowand Codex getsmodel_reasoning_effort="low". - The local speech model runs a one-second practice transcription of quiet noise after it loads, after you switch models and 3 seconds after wake or unlock. A dictation started meanwhile cancels it. On the Mac it was tested on, the gain was within measurement noise.
make shippackages DMGs with macOS'sdiskutil imageinstead of thecreate-dmgtool, so it no longer printshdiutildeprecation warnings on macOS 27 and no longer needsbrew install create-dmg. The disk image is APFS instead of HFS+ and keeps the same window layout. Packaging DMGs now needs macOS 27 or later; the DMGs still open on macOS 14.
Fixed
- Light cleanup now turns "I live in Delhi no sorry I live in Ahmedabad" into "I live in Ahmedabad". A "no sorry" or "sorry, no" counts as a correction even when the speech model writes no commas, and a correction can restart on "I" or "we" when the word after it repeats too. "Should we go, no, we should stay" and "I said no sorry I can't come" stay as you said them.
- Light cleanup no longer deletes words when a "no" or "I mean" only looks like a correction. "We read it in the last book, no, the whole thing with the houses was different" lost "in the last book". A correction now has to end about as soon as the words it replaces, as "Let's meet at 5, no, at 6 pm" does.
- Light cleanup keeps repeats said for emphasis. "Drinking more and more and more" lost its last "and more", and "again and again and again" lost one "and again", as if they were stutters. "And then and then we shipped it" still becomes "and then we shipped it".
- The Parakeet phrase booster no longer replaces correctly spelled words with a vocabulary term that sounds like them. With Claude in Vocabulary it wrote "Claude" when you said "cloud", and on 597 recorded Indian English segments with 20 terms nobody said, it made 9 of those segments worse with Parakeet TDT v3 and 10 with Parakeet Ultra. A word that is spelled correctly and looks clearly different from every term now stays, which made none of them worse, while a misheard name such as "Rivali" still becomes "Diwali". Spelling is checked with the macOS English dictionary.
- Meeting notes stick to what was said. The notes model listed things that already happened in the meeting, such as introductions, as action items, gave Me tasks Me never took on, and wrote "No explicit decisions were made" as a decision. Notes now count only what someone agreed or took on, and drop items that only say there were none, while keeping real ones such as "None of the current vendors will be renewed". On seven recorded meetings, three notes each, an AI judge reading the transcripts found 1.3 unsupported claims per set of notes against 4.1 before, and 0.2 filler items against 1.4.
- Long meetings no longer fail with "The model couldn't condense this meeting enough" on Apple Intelligence. Asked to shorten the transcript, the model copied its lines back, so it never got shorter; 2 of 3 runs on a 10-minute four-person meeting failed. It now writes a short summary of each part in its own words.
- Learn from my corrections now notices a fix even when you also changed text elsewhere in the same field, and in web fields that report spaces and line breaks differently from the pasted text. Before, either case made it miss the fix.
- Fixing a dictation in History no longer moves the cursor to the end after every key you press. Text typed in the middle now goes where the cursor is.
v0.1.4Oct 10, 2026
Added
- English dictation writes spoken numbers as digits with every speech model, including Parakeet models that write numbers as words. "Two hundred and fifty thousand" becomes 250,000, "one hundred or two hundred" becomes "100 or 200", "two point five million" becomes "2.5 million" and "zero point one point one" becomes 0.1.1. A lone small number ("give me five minutes"), "a hundred times", ordinals, years said as pairs ("nineteen ninety nine") and digit-by-digit phone numbers stay as spoken.
- Spoken numbers understand Indian English "lakh" and "crore": "one lakh" becomes "1 lakh", "two lakh fifty thousand" becomes 2,50,000 and "two point five crore" becomes "2.5 crore". "Twelve hundred" becomes 1200, and a phrase that is not one number keeps its readable part as digits, so "ten thousand hundred followers" becomes "10,000 hundred followers".
- English dictation writes spoken symbols. "At the rate" or "at sign" writes @ ("ping at the rate KV" becomes "ping @KV", "name at the rate gmail dot com" becomes name@gmail.com), "hashtag launch" becomes #launch, "issue hash forty two" becomes "issue #42" and "ten percent" or "ten percentage" becomes 10%. Saying only "at the rate", "hash", "hashtag", "percent" or "percentage" writes just the symbol. Phrases like "hash out a plan" and "what percentage of users" stay as words. An address the model hears as "gmail. com" or "at the rategmail.com" is joined back into one address.
- Vocabulary can fix a voice command the speech model mishears in your accent. Add what it wrote as the spoken phrase and the command as "Write as", such as "Kocia Mark" written as "question mark", "Coma" as "comma", "at the Red" as "at the rate" or "Hai fun" as "hyphen". These entries run before voice commands, so the dictation gets "?", a comma, "@" or "-". They are not sent to the phrase booster, because boosting command words dropped neighbouring words in testing.
- Settings has a Vocabulary row under Microphone and language that explains how your word list helps and opens it. With a Parakeet model it tells you when the phrase booster isn't downloaded yet, since Parakeet needs it to use your words while recognizing speech.
- When a call starts in Zoom, Google Meet, Microsoft Teams, a Slack huddle, Webex, FaceTime, Discord, WhatsApp, Signal, Telegram, Whereby or Jitsi Meet, the pill offers Start notetaker. It asks once per call and the offer goes away after 30 seconds. Calls in a browser are recognized by their window title, which needs Accessibility. Turn it off with Detect meetings in Settings or from the offer's menu. It needs macOS 14.2 or later.
- Settings has an Appearance section with a Theme choice of System, Light or Dark. BetterWispr now starts in Dark; choose System to follow your Mac as before.
- "Exclamatory mark" writes "!", and saying only "hyphen", "colon" or "dash" writes just that mark.
- Say "write an email to Sarah saying I will be on holiday tomorrow, thanks, best regards, Kartik" and BetterWispr types it laid out as an email: "Hi Sarah," then the body, then "Best regards," and your name on their own lines. "Draft" and "compose" work too, as do "that", "telling her" and "letting him know". It works on this Mac without a notes model, in English or Auto, and skips the tone and the Medium edit so the layout stays. If the speech model misspells your name, add it to Vocabulary.
- Style lists the apps you recently dictated into under each context, so you can see which tone each one gets. T3 Code, Chrome, Brave and other apps outside the lists use Other apps, which is Formal until you change it.
- Smoother motion. Capsule cards and pills grow out of the pill as they fade in, the dashboard cross-fades between pages, the toast slides back up as it leaves and Style reveals the notes model picker smoothly. With Reduce Motion on, these are plain fades.
- Spoken language lists the languages the selected model understands. Parakeet TDT v3 and Parakeet Ultra now offer all 25 of their European languages, such as Polish, Dutch, Ukrainian and Greek, instead of a fixed list of 11 that included five they cannot recognize (Hindi, Japanese, Korean, Chinese and Arabic). A language you chose earlier that the model does not support stays visible and is marked as not supported.
BetterWisprCLI --bench MANIFEST.jsonltranscribes a list of clips with one loaded model and prints each transcript with its recognition time, optionally after the app's Light or Medium cleanup or with a--vocabularylist for the phrase booster.scripts/evaluate_transcripts.py --group languagescores each language separately, andscripts/fetch_bench_corpus.pydownloads a small LibriSpeech and Multilingual LibriSpeech test set to run them on, or with--accentsconversational Indian, Nigerian and US English from EdAcc.
Changed
- The dashboard opens on a new Home page. It shows your shortcut and a button to start dictating, your words, speaking time, pace and streak, your most recent dictations, and whether your speech model, microphone and pasting are ready, each one a click away from fixing it. The sidebar groups pages under Personalize and App.
- Settings is laid out as cards (Shortcut, Microphone and language, After you speak, App, Permissions and Voice commands), and each option has a short line explaining what it does.
- Style shows what each cleanup level does to a sample sentence as you switch levels. Each app context has its own card with an icon, its tone choices and a preview of the tone, and the apps it covers are listed underneath.
- While a dictation is being transcribed, the pill shows a spinner instead of the moving bars.
- The notetaker window is calmer. The date and speech model share one line under the title, the tabs are plain text with an underline that slides to the open tab, and the transcript tools sit on the page instead of in a gray box. While recording, the bottom of the window is one bar with the timer, live Me and Them waveforms, a microphone menu and a red Stop button, with the consent reminder underneath. Before the first words arrive, the transcript says it is listening instead of showing an empty box.
- Recording waveforms in the pill, the notetaker and Home now scroll from right to left with your voice. Each bar is a moment of sound, louder syllables stand taller and pauses drop flat, instead of a fixed shape that wobbled at one height.
- Automatic microphone follows Sound settings again, so AirPods and other Bluetooth headsets are used when macOS picks them, with the lid open or closed. A chosen microphone that is disconnected now falls back to the same macOS default instead of skipping a connected headset. Recording from a Bluetooth headset's microphone still drops its playback to call quality until you stop; choose your Mac's own microphone in Settings to avoid that.
Fixed
- A new version no longer arrives silently. When BetterWispr finds an update while you are working in another app, a toast offers Update and the menu bar shows "Install BetterWispr" with the new version, instead of Sparkle's window opening behind your other windows. With automatic installs on, a downloaded update shows a Restart toast and menu item, because the update otherwise waited until you quit BetterWispr.
- With Spoken language set to Detect automatically and words in Vocabulary, Parakeet no longer rewrites other languages into your vocabulary terms. The phrase booster understands only English, and it turned German "meine" into "Mike". It now runs automatically only when the transcript reads as English.
- The Parakeet phrase booster no longer swaps everyday words for your vocabulary. It turned "not like" into "Notion" and "Thank" into "Thai", and on 300 recorded conversation clips with 20 unspoken terms it changed 44 of them and raised the error rate from 16.2% to 17.4%. A word must now sound closer to a listed term before it is replaced, which changed 5 of those clips. It also no longer drops the word next to a term, as when "seven Harry Potter movies" became "Harry Potter movies".
- Light cleanup resolves a correction after a lone "sorry" that restarts on in, at, on, from, near, into, by or with. "I live in Ahmedabad sorry in Delhi" becomes "I live in Delhi" and "meet at 5 sorry at 6" becomes "meet at 6". Apologies such as "sorry for the wait", "so sorry in advance" and "sorry, the train was late" stay.
- Cleanup no longer drops a repeated number word as a stutter, so "twenty twenty five" and "fifty fifty" keep both words.
- Meeting notes no longer lag while recording. The Me and Them meters used to redraw the whole notetaker window and the floating pill on every audio buffer, about 140 times a second. Now the levels update at most 30 times a second and only the meters and the clock redraw.
- The pill no longer keeps spinning for about three quarters of a second after your words are pasted, and the next dictation can start right away. BetterWispr used to wait in place before putting your previous clipboard back. It now restores the clipboard in the background, and when you dictate again before that happens, the clipboard you had before the first dictation is the one that comes back.
- Tapping the notetaker pill while a meeting is still starting no longer silently cancels it. A tap stops a meeting only once it is recording; Cancel in the notetaker window, the pill's menu and the menu bar still cancel a meeting that is starting.
- Meeting notes no longer wait up to 30 seconds to show what you said, and no longer cut your sentences in half. BetterWispr ended a microphone chunk at a pause only when the level fell below a fixed threshold, which steady room or AirPods noise never did, so chunks ran to 30 seconds. A pause now counts when the level falls back to the background noise of the last 10 seconds, which also ignores the short silent gaps AirPods leave in their signal, and chunks are 3 to 20 seconds long.
- Meeting transcripts no longer repeat the other person's words under Me when your microphone hears them through your speakers. Repeated audio used to be hidden only when a whole Me entry matched them. Now the matching words are removed and what you said around them stays, so "Sure. We should ship the new flow before the end of the month. I agree completely." becomes "Sure. I agree completely." when the other person said the middle sentence at the same time. A Me entry with fewer than three of your own words left is hidden. Show repeated microphone audio still shows the original.
- The other person's lines in a meeting transcript now sit at the right time next to yours. BetterWispr timed their audio by counting what it received, but macOS only sends it while something plays, so their entries were placed seconds early (3.7 seconds in a test) and some of their words picked up by your microphone were not removed from your entries. Each chunk now takes the time its audio was captured.
- Meeting notes start recording as soon as you start them, while the speech model loads. The first words of a meeting used to be lost and the first entry appeared 3 and 11 seconds after the line ended in two tests. What you record while the model loads is transcribed once it is ready, and in testing the first entry appeared about 1 second after the line.
- The other person's last line now appears about a second after they stop talking. macOS sends their audio only while something plays, so when their app went quiet the line waited for the next sound, 6.2 seconds once in a test.
- Repeated microphone audio is also removed when the two transcripts spell a word differently, such as "campaign" and "campage", and a repeated sentence no longer leaves its first word behind in your entry. Words the speech model invents from the repeated audio, rather than repeats, can still remain.
v0.1.3Oct 8, 2026
Fixed
- Videos and music in AirPods and other Bluetooth headphones keep their full sound quality while you dictate. Recording from a Bluetooth headset's microphone switches its playback to call quality, so Automatic now uses your Mac's built-in microphone instead while the lid is open. With the lid closed, or on a Mac without a microphone, Automatic follows Sound settings. If you chose your headphones in Settings, they are still used, so switch Microphone to Automatic to keep full sound quality.
- Successful dictation plays only the recording-stop sound; insertion and copying no longer play a second completion sound. Notetaker remains silent.
- Microphone callbacks explicitly avoid inheriting MainActor isolation, preventing the capture-start crash reported on the audio worker queue.
- Quick taps in Hold to talk start hands-free dictation with compact cancel and finish controls. Cancellation offers Undo with a countdown; successful insertion returns to idle without a completion message. The notetaker pill has more waveform width while keeping its compact height.
- Dictation feedback uses only locally installed sounds. The earlier sound packs and all fallback cues have been removed.
- Dictation and meeting notes wait for microphone audio before reporting readiness and finish the in-flight audio buffer before transcription. Both use input-only Core Audio capture with 20 ms buffers instead of the audio engine tap's larger batches. Dictation starts capturing without waiting for the speech model to prepare. Bluetooth startup runs off the UI thread, and a running input queue handles Bluetooth rate changes without a second restart.
- Quiet, brief speech is checked in small windows so larger Bluetooth audio buffers do not dilute it below the silence threshold. Waveforms respond after digital silence from noise-cancelling headsets and show the latest delivered level without replaying a buffered delay.
- Dictation and meeting notes recover microphone capture after AirPods and other external microphones change the device's audio configuration. Device changes are allowed to settle, and a failed restart is retried before reporting an error.
- Automatic microphone selection follows device changes during dictation as well as meetings. Audio recorded before a switch is kept, including when the new microphone uses a different sample rate or channel count; meeting system audio continues independently.
- Changing the microphone in Settings also applies to a dictation already in progress. Pending device changes are cancelled when recording stops or is cancelled.
v0.1.2Oct 8, 2026
Fixed
- Downloading a speech model or the phrase booster no longer quits BetterWispr. In 0.1.1, every Parakeet and Whisper download, from Models, the welcome guide or Vocabulary, closed the app as soon as it started, so no model could be installed.
- Your Mac no longer falls asleep from inactivity partway through a model download.
- If the phrase booster fails to download alongside a Parakeet model, the progress bar still reaches 100% instead of stopping short and jumping to "Setting up on this Mac".
- Cancelling a download and choosing Download again right away waits for the cancelled one to stop, instead of loading the same model twice at once.
- A Whisper download that Hugging Face refuses now says so in plain words instead of showing "ArgmaxCore.Downloader.DownloadError error 1".
v0.1.1Oct 8, 2026
Added
- More local models in Models: Parakeet Ultra (a further-trained TDT v3 with punctuation), Parakeet TDT-CTC 110M (a small English model), Parakeet Japanese, Whisper Large v3 Turbo (compressed) and Whisper Small. Each downloads only when you choose Download.
- Help > Show Welcome Guide replays the welcome guide.
BetterWisprCLI --download-modelshows install progress.- Style page with a writing tone for each kind of app. Personal messages offer Formal, Casual and Very casual. Work messages, Email and Other offer Formal, Casual and Excited. Casual drops most commas and the final period, Very casual also lowercases sentence starts, and Excited ends the last sentence with an exclamation mark. BetterWispr picks the tone from the app you dictate into, so Slack gets your work style and Mail your email style, while browsers and AI apps use Other. Styles change English dictation only, run on your Mac, and default to Formal, which leaves your text as it was.
- Auto cleanup on the Style page. None keeps your words as recognized, still applying voice commands and Vocabulary. Light, the default and the previous behavior, removes filler words like "um" and "you know" and stutters. Medium also edits English dictation for clarity and conciseness with the notes model chosen in Models (Apple Intelligence, Ollama, Claude Code, Codex or an API connection), so it works with whichever one you already use for meeting notes. Apple Intelligence and Ollama run on your Mac. Claude Code, Codex and API connections send the dictation to that provider and add a few seconds. Other languages, very short dictations and edits that fail, take too long, answer the dictation or add new content keep the Light result. The Style page uses standard macOS grouped settings, and choosing Medium shows the notes model it will use.
- Insights page with your words per minute compared with average typing speed, words cleaned up, vocabulary fixes, total words dictated, the kinds of apps you dictate into, and a streak calendar with your current and longest streak. Insights are computed on your Mac from saved history, and the Share button offers a one-line summary through the macOS share menu.
- Use Original in History, under Original transcription, puts back what you said when cleanup or a style changed it.
Changed
- Parakeet installs fetch the model in one pass instead of one pass per file, and the phrase booster downloads alongside it by default. Model sizes in Models include the booster, which downloads once and is shared by every Parakeet model. If the booster download fails, the model still installs and Vocabulary offers it again.
- Model downloads run separately from dictation. You can dictate with the current model while another downloads, and When it finishes, BetterWispr switches to the new model, unless you picked a different one meanwhile or are mid-dictation.
- Download progress only moves forward and covers the whole install, from the download to setup on this Mac. Models, the welcome guide and Vocabulary offer Cancel during a download and Retry after a failure.
- The welcome guide's completion is saved with your workspace and versioned, so a reset workspace shows the guide again. Quitting midway reopens the guide on the same step. The 0.1.0 preference flag is migrated once and then removed.
- New dictations save the app they were sent to and how many words Vocabulary respelled, for Insights. Dictations saved by 0.1.0 load unchanged and count toward every total except app usage.
- Adding a Smallest AI connection now asks for the spoken language in the connection sheet and saves it to Settings, since Smallest AI can't detect the language itself. Before, you learned that only from an error after you started dictating. The picker lists only languages Smallest AI accepts. If Spoken language is later set to one it doesn't, such as Arabic, BetterWispr stops before recording and says which setting to change.
- Light cleanup also drops a spoken correction when the fix repeats a word you just said, so "I want to go to Pune, sorry, no, to Delhi" becomes "I want to go to Delhi". It needs "no" or "I mean" after a comma, so a plain "sorry" apology, "yes to this, no to that" and corrections that share no word ("Pune, sorry, no, Delhi") stay as you said them.
Fixed
- The welcome guide's speech model download no longer stops when Escape, the capsule, the menu or a shortcut release cancels dictation, and a failed download now shows its error in the guide.
- Parakeet Japanese never applies the English phrase booster.
- Insights counts the words the Parakeet phrase booster spells from your Vocabulary while it recognizes speech. Before, those dictations came out right but counted no vocabulary fixes.
- A dictation you switch away from before it finishes is no longer easy to miss. When BetterWispr can't paste into the app you started in, because you moved to another app or it quit, the capsule opens a card with your text and a Copy button at the top right. A ring around the button counts down five seconds, then the card closes itself. The text is also on the clipboard, so ⌘V works too. Before, only a brief "Copied" toast at the top of the screen said where it went.
- Learn from corrections catches a fix you send right away, as in chat apps that clear the field on send. BetterWispr now checks the field every half second and learns from text that stayed unchanged for half a second before the field emptied. A fix you leave in place is still learned after two seconds without changes.
- Learn from corrections works in Electron apps such as Claude and T3 Code. These apps hide their text fields from Accessibility until an assistive app asks for them, so BetterWispr could never see your fix there. When learning is on and the field isn't readable after a paste, BetterWispr turns on that app's Accessibility support and checks again for up to five seconds.
v0.1.0Oct 8, 2026
Added
- Meeting summaries with signed-in Claude Code and Codex subscriptions, or a custom OpenAI-compatible chat-completions endpoint and API key stored in macOS Keychain. Choose the notes provider independently of speech recognition; Apple Intelligence remains the default, with no automatic cloud fallback.
- Claude Code and Codex model pickers populated by the installed CLIs, including exact model IDs, refresh and custom model entry. The chosen model is saved and identified on generated summaries.
- A notes-model test in Models that uses a short synthetic sample, with cancellation and actionable connection errors.
- Meeting notes and titles with Ollama. Choose an installed Ollama model under Notes model in the Notetaker settings menu, and BetterWispr writes the title, summary, key points, decisions and action items with it on this Mac. Apple Intelligence stays the default, BetterWispr never switches models on its own, and cloud and embedding models are not offered.
- Voice commands for English and Auto language: "comma", "question mark", "full stop", "add a period", "new line", "new paragraph", "scratch that" and "sorry, remove that" delete the last sentence, and "at the rate KV" or "at sign KV" types "@KV". Commands are rule-based and run on your Mac.
- Learn from corrections: fix a misheard word in History with the pencil, or in the text field within 30 seconds of a paste, and BetterWispr adds it to Vocabulary with a Learned badge and a toast. Turn it off in Settings.
- Native macOS dictation app with a global ⌥ Space shortcut, paste into the focused app and transcript history.
- Local speech models: Apple speech, Parakeet TDT v3 (25 European languages) and v2 (English) through FluidAudio, and Whisper large-v3 turbo through WhisperKit.
- Compact recording overlay with an idle pill, a hover tooltip showing the chosen shortcut and a dictate button.
- Welcome guide on first launch. It fills the main window, walks through permissions, the recommended speech model and a first dictation, and runs only once.
- Disk image window with BetterWispr and Applications side by side on a warm paper background, a hand-drawn arrow between them and the hint "Drag to Applications".
scripts/dmg-background.swiftrenders it at 1x and 2x so it stays sharp on Retina displays, and Finder's labels read in Light and Dark Mode. - Personal vocabulary with spelling hints and whole-phrase replacements, plus local model management.
- Optional local phrase booster for Parakeet vocabulary. Download it explicitly from the Vocabulary page to help recognize names and technical terms. The English-only booster runs with English or automatic language selection, preserves punctuation around replaced words, and loads installed assets offline.
- Hold-to-talk dictation, now the default: hold ⌥ Space while speaking and release to finish. "Press to toggle" keeps the previous behavior and is set in Settings.
- Change the dictation shortcut in Settings. Click the shortcut, press a new combination that includes ⌘, ⌥ or ⌃ or an F-key such as F5, or reset it to ⌥ Space. The recorder shows the keys you hold and says why a key it refuses can’t be used. Hints, menus and error cards show the shortcut you chose.
- "Copy to clipboard" setting, off by default. When on, each dictation stays on the clipboard after it is pasted. When off, BetterWispr restores the previous clipboard, and with paste also off it leaves the clipboard alone. A blocked paste still copies so the text is never lost.
- Settings for spoken language, input sensitivity, automatic paste, local transcript history, the floating capsule and opening at login.
- Error card above the capsule with a shake and red outline for taps, early releases, missing speech, blocked paste and other failures. It dismisses itself after six seconds and respects Reduce Motion.
- Toast confirmations for copy and delete, adapted from BetterShot's toast panel.
- Menu bar icon switches to a microphone while recording.
- App icon and in-app logo: a hand-drawn "w" on a light tile, shown on a dark tile in Dark Mode.
scripts/build-app.shrenders the icon from the same SwiftUI view. - Dedicated About page with the app logo, version, source code and issue links, and the author's X profile.
- In-app updates from GitHub Releases through Sparkle 2. Check from About or the app menu, and choose whether BetterWispr checks daily and installs new versions on its own. Updates install only when their EdDSA signature matches the app's public key.
- Meeting notes. Start a meeting from the Meetings page or the menu bar to record your microphone as "Me" and other apps' audio as "Them" (macOS 14.2 or later), follow a live transcript, and write your own notes. When it ends, Apple Intelligence on a supported Mac running macOS 26 or later writes a summary, key points, decisions and action items locally. Without Apple Intelligence the transcript and your notes are still saved. Search saved meetings, mark action items complete and copy meeting notes and transcripts as Markdown. With a built-in speech provider, audio stays on the Mac; API providers require explicit selection.
- Website with a typing and dictation demo, feature overview, comparison pages, download links, privacy and terms pages, and a dedicated changelog rendered from this file. Pages include static content and social previews.
BetterWisprCLIfor local model downloads and file transcription, defaulting toparakeet-v3. Pass comma-separated spelling hints with--vocabularyor explicitly install the Parakeet booster with--download-phrase-booster.- Local transcript evaluation script reporting word error rate, character error rate and false speech on silent clips, with an accuracy evaluation plan.
make devfor local testing andmake shipfor separate Developer ID-signed, notarized and stapled Apple Silicon and Intel DMGs, with a Sparkle update feed that selects the matching architecture. Notarization credentials are stored once in Keychain withmake setup-notary.VERSIONfile and this changelog. The app bundle version is stamped fromVERSIONat build time.
Changed
- Meeting summaries combine the transcript and all of My thoughts, including longer notes. Edited thoughts show when the summary needs updating, with an Update summary button.
- The menu bar menu shows whether BetterWispr is ready, listening or transcribing, switches the speech model and microphone, and opens Settings and update checks. Show capsule and Show welcome guide are gone.
- Start Notetaker on the dashboard now docks the meeting card beside your call, the same as starting from the capsule or the menu bar.
- The dashboard no longer opens the live transcript while you record. It keeps the list of notes and opens the finished note once recording and notes are done.
- The app bundle's
CFBundleVersionnow followsVERSIONinstead of a fixed1, so Sparkle can tell releases apart. - The capsule no longer shows the recognized text above the waveform while you dictate or while it transcribes.
- A successful paste no longer shows a "Pasted" toast. Copy-only results still confirm with a toast.
- Recording capsule redesign. Holding ⌥ Space shows only a waveform. In toggle mode the capsule shows cancel, the waveform and a stop mark, and a click anywhere on it finishes dictation. A failed session collapses back to the idle pill with a red outline under the error card.
- Waveform that fits your voice. It learns the room's noise floor and your recent loudness, so quiet voices fill the bars and steady background noise stays flat. Bars rise fast and fall slowly, follow the voice in steps of about 20 ms instead of jumping with each 100 ms audio buffer, and a travelling wave shows while transcribing. Reduce Motion keeps the bars still.
Fixed
- Escape now cancels an active dictation from the dashboard without adding a history entry. DMG packaging retries Finder's temporary busy error, and release preflight distinguishes Apple agreement errors from invalid credentials.
- Hide matching long microphone echoes of nearby system-audio transcript entries in the transcript, export and summarization input. Original entries remain saved and can be revealed with Show repeated microphone audio.
- Cancelled or failed summary generation preserves the saved meeting and previous summary. Oversized notes fail explicitly if they cannot be condensed, instead of silently losing the end of the source.
- An open note now switches to its Summary tab as soon as the notes are written, instead of staying on the transcript.
- Paste into other apps stopped after every rebuild because ad-hoc signing changed the app's code requirement and macOS revoked Accessibility.
scripts/build-app.shnow signs with a stable identity, and a blocked paste shows an Allow button instead of silently copying. - English transcript cleanup removes "uh", "um", "er", "hmm", "mm" and their supported spellings, set-off "you know", and unpunctuated back-to-back repeats such as "which you which you" before vocabulary replacements. It preserves "mm" after a numeric value, "you know?", punctuated repeats, numbers and common doubles such as "that that" and "long long". Language detection ignores fillers; other languages are left as spoken, and history keeps the raw recognizer output.
Your next thought, in your own words
Give your keyboard a break.
Local speech models. The apps you already use. A little more room to think.
“Just tried it out as I was looking for something that does this exact thing well (macbook). So far I dictated about 5 different things in different contexts, it worked without a single error!”
macOS 14+ · Apple Silicon recommended · No account in the app