Changelog

What’s new, what’s better

175 changes in 21 releases, from the 1.0 launch on 27 September to version 1.5.6. Every release has its own entry, dated the day it shipped. Already installed? ZenVoice updates itself.

1.5.6
12 November 2026
Latest
Snippets you can trust, lists and email addresses written the way you said them, and a tidier model lineup.
New
Snippets can have aliases: other phrases, one per line, that insert the same content.
Snippets can be limited to the apps you choose. An app-specific snippet can share a trigger with a global one and wins in that app.
Turn on “Replace {date} with today’s date” to have a snippet insert the current date when it expands.
Search your snippets, duplicate one into a new draft, copy its content, and undo a deletion. Starter examples help you write your first ones.
Export all your snippets to a file and import them on another Mac. Imports add to your snippets and never replace them.
The Models card for Qwen3 1.7B now shows its measured accuracy and speed.
Better
Snippets match more reliably. Exact and longer phrases win, short words must match exactly, and longer phrases forgive a small slip. Inserted content can no longer expand itself or set off another snippet.
Snippets settings has a cleaner layout, and Try it shows the full result and which snippets matched.
Spoken lists become numbered lists more often, and when a list is ambiguous, Qwen3 1.7B can help decide where each item ends. Clear cases are handled the same way by Qwen3, Sumi and Fude.
Spoken email addresses are written out: “john dot smith at gmail dot com” becomes john.smith@gmail.com.
Punctuation words you mean as words stay words, so “the dot product” or “a comma” is no longer turned into a symbol.
The notch HUD has smoother upper curves, and its recording strip is wider.
Fixed
ZenVoice releases its speech models cleanly when you quit.
Removed
Kanso Koi 1.0 is retired from setup and the Models page. If you already use it, it stays selected and keeps working, and the command line tool can still install, use and remove it.
Hinglish Prime is retired from new downloads, setup and recommendations. Existing Prime installations remain usable and can be removed in Models.
Qwen 3.5 is retired from setup and the Models page. The command line tool can still use it.
1.5.5
9 November 2026
Everything from 1.5.1 to 1.5.4 in one update: choose the pill's size, Parakeet Unified on the Neural Engine, a redesigned Insights page, and a quicker recording pill.
New
Choose the floating pill's size. In Settings → Recording HUD, Floating panel now has Width and Height sliders, from the size of Hex's bar (112 × 26 pt) up to the full 1.5.0 pill (268 × 44 pt). The pill appears in its corner while you drag, so you can see the size you're choosing. The size applies while the pill shows your voice as dots, and NO SPEECH DETECTED fits itself to it: on one line, on two, or as NO SPEECH on the smallest pills. Live words, notices and other errors still use the full pill.
Parakeet Unified · Neural Engine, a second way to run Parakeet Unified. It uses your Mac's Neural Engine instead of the GPU, so dictation stays quick while games, video or other apps keep the GPU busy: in testing, the wait after you stop stayed near 0.04 s where the GPU version took 0.47 s. It makes about as many mistakes as Parakeet Unified (7.8% of words against 7.7% on ZenVoice's test recordings), gives the same Dictionary fixes and language checks, and is a 0.7 GB download. It writes your words when you stop speaking, not while you speak. After the download, ZenVoice prepares it for your Mac once, which takes up to a minute. Find it on the Models page, next to Parakeet Unified.
Insights: Time saved, how long typing your words at 40 words per minute would have taken, minus the time you spent speaking.
Insights: When you dictate, your busiest days and hours at a glance.
Insights: Personal records, your best day, longest dictation, fastest dictation and longest streak.
Insights: Next milestone, how close you are to your next word count, and when you'll reach it at your pace.
Insights: Fixes per 100 words, which falls as ZenVoice learns your words.
Better
When a dictation hears nothing, the recording bar no longer swaps to a warning. Its dots turn red and spell out NO SPEECH DETECTED in capitals made of the same dots, in English whatever language ZenVoice is in, then the bar closes after about a second. Your chosen error sound still plays.
The recording bar dismisses sooner after a successful dictation or when no speech is detected.
The menu bar panel is smaller and matches the app's dark look, with Record and Paste side by side. C1, C2 and C3 copy your three most recent dictations, and a speech model picker shows the model in use and lets you switch. History shows each app's icon, words and words per minute, a 14-day chart with every day labelled, and your total for the week.
Insights has a new look in ZenVoice's colours. It opens with a one-line recap, then four cards: Words, Time saved, Speed and Dictations, each with its change since the period before.
Growth always shows now. Days without dictations count as zero, and while your history is shorter than the range you pick, Insights compares your latest days with your first ones.
Choose bars or lines for every Insights chart. Lines run edge to edge with a soft fill.
The streak calendar uses blue, like the rest of the page.
Privacy and Share sit at the top of Insights, and the privacy note is shorter.
The ZenVoice window opens a little wider, at 935 × 720.
"Preparing for this Mac…" now says it takes up to a minute, since Parakeet Unified · Neural Engine takes longer to prepare than Parakeet V3 · Neural Engine.
Fixed
The floating pill shows at once when you start a dictation right after the last one. It used to wait up to a second and a quarter while ZenVoice updated History and Voice profile in the background.
Dictation sounds no longer hold up the recording pill. Each sound waited for your speakers or headphones to wake before the pill could appear, which on Bluetooth could take a second or more.
Closing the menu bar panel no longer leaves ZenVoice using about 12% of a CPU core until you quit.
Lower idle CPU use by pausing animated settings previews and onboarding animations while their window is hidden. The recording pill's dots also pause once the pill leaves the screen.
The recording bar no longer waits for the focused app's Accessibility response before appearing, keeping shortcuts and animations responsive.
Removed
The Patterns card on Insights. When you dictate replaces it.
1.5.0
7 November 2026
Kanso Sumi and Fude format in milliseconds, Parakeet V3 runs on the Neural Engine, and the end of a dictation is no longer cut off.
New
Kanso Sumi and Kanso Fude, new formatting models for Polished English. They mark each word's punctuation and capitals and drop fillers, repeats and take-backs, but never write a word you did not say. On ZenVoice's 500 test dictations Sumi makes fewer mistakes than Kanso Satori 1.0 (4.0% of words against 4.7%) and formats a dictation in about 6 ms instead of 350 ms, using about 360 MB of memory. Fude, built on Gemma 3 270M, removes fewer real words from long dictations. Download either on the Models page.
Parakeet V3 · Neural Engine, a second way to run Parakeet V3. It uses your Mac's Neural Engine instead of the GPU, so dictation stays quick while games, video or other apps keep the GPU busy: in testing, the wait after you stop stayed near 0.06 s where the GPU version took 0.7 s. It gives the same words, Dictionary fixes and language checks as Parakeet V3. After the download, ZenVoice prepares it for your Mac once, which takes about 20 seconds. Find it on the Models page, next to Parakeet V3.
zenvoice formatting lists, installs, chooses and removes formatting models from Terminal, and zenvoice doctor reports the one in use.
Better
Kanso Satori 1.0 and Kanso Koi 0.9 are retired from the Models page. Their downloads keep working for anyone who chose them, and zenvoice formatting install and use can still set them up. A retired model shows in the app, marked Legacy, only while it is the one chosen.
After a macOS update, ZenVoice prepares Parakeet V3 · Neural Engine again when it starts, and the menu bar says “Preparing your speech engine…”. If you dictate meanwhile, the bar shows “Getting ready…” and your words appear as soon as it's done.
Fixed
The end of a dictation is no longer cut off. On macOS 27.2, up to the last 64 ms of every recording was lost with every downloaded speech model, including Parakeet, Nemotron, Whisper and Hinglish Prime, which could drop or garble a short final word.
1.4.0
5 November 2026
Dictation sounds, the Zen Dots HUD and notch island, a refreshed setup, and Parakeet fixed and faster on M5.
New
Dictation sounds. ZenVoice plays a short tone when dictation starts, stops, finishes and fails. Choose a set on the Dictation page: ZenVoice Signature, a softly struck meditation bowl made for ZenVoice, or Classic, Click, Soft tone, Wood and Warm synth, from Hex. All six play at the same loudness, so the volume suits each one. Choosing a set plays it once. Turn Play sounds off to keep dictation silent. The finished tone plays only when text arrives two seconds or more after you stop, so a quick dictation doesn't chime twice.
Better
Zen Dots: the dictation bar, the Live Preview pill and the notch now draw your voice on a plate of glowing dots that fills them edge to edge, into the rounded ends, at 60 frames a second. It arrives with a glint and a line that shoots out to both ends. Choose how your voice is drawn under Bar animation on the Dictation page: Strands, two threads that braid into knots as you speak; Loops, Hex's own animation, two mirrored threads that close into loops; Waveform, your voice drifting past from right to left; Spectrum, its pitches mirrored from the middle; or Pulse, waves spreading out with each word. While ZenVoice transcribes, bands of light sweep across like raked sand. With Reduce Motion on, the dots hold still.
The floating HUD now comes and goes as Hex's bar does. It appears small and blurred and springs out to its width. Without live words, the dots carry the whole dictation: the pill narrows while ZenVoice transcribes and inserts (with Loops, a cooler helix with pulses running along it), then a wipe clears it from the middle and it blurs and fades away, instead of showing “Transcribing…” and “Inserted”.
On a Mac with a notch, the notch HUD is now Hex's island: a black shape that grows out of the notch on a spring, 28 pt past it on each side, holding a field of dots below the camera, and shrinks back into the notch when done. It meets the top of the screen in soft curves, and its words, notices and cards open inside it the same way.
The cancel and finish buttons are gone from the recording bar, pill and notch. Press your dictation shortcut to finish or Esc to cancel, as before; clicks while you dictate now go to the app underneath.
Setup has a fresh look: centered headings, native artwork for every model, and layouts that fit the window. Welcome shows the models circling ZenVoice, Formatting shows all four options at once, the HUD picker matches the new notch and floating pill, and the last step shows your own dictation shortcut.
Fixed
On Macs with an M5 chip, Parakeet no longer gets stuck repeating a word (“tamb tamb tamb…”). Parakeet Ultra, Parakeet TDT V3 and V2, and Nemotron now use the GPU's faster tensor path, about 20% faster on M5.
1.3.3
2 November 2026
A compact Share Highlights card, sooner Insights comparisons and a clearer License page.
Better
Share Highlights is now one compact card for the selected Insights range: words dictated with the change against the prior period, speaking speed, day streak and apps used, your five most used apps, and an activity chart. Choose Graphite or Paper and Bars or Line; images export at 2400 × 1260.
Insights compares sooner. Instead of "Incomplete history; comparison unavailable", a new history is compared against your saved dictations from the same number of earlier days, labelled as a saved trend, or says "Keep dictating to build your comparison" until there's enough.
License shows your license at a glance: who it's licensed to, the plan, the key with its middle hidden, how many Macs it covers, when you bought it and which updates it includes. Moving to another Mac, upgrading and managing your licenses each have their own place.
Each page's header shows the same icon as its sidebar row, and the header is smaller so more of the page fits.
Kanso Hinglish Prime has its own icon in Models and setup.
1.3.2
29 October 2026
A new icon, dictation that lands in every app, and fixes for silence, times and Dictionary words.
New
Per app, in Settings → Insertions: choose Auto, Paste, Type or Copy only for any app. Type enters the text key by key on your keyboard layout, for a field that refuses paste, and leaves your clipboard alone. Auto types into remote desktops and virtual machines and pastes everywhere else.
Better
A new ZenVoice icon, Graphite & Pearl, in the Dock, the menu bar, setup and About.
New icons for every page in the sidebar, drawn for the dark sidebar and sized to fill it. They keep their colours when a page is selected.
Models and setup show each model maker's logo, now with crisper artwork, and setup shows the maker beside each recommended model.
In T3 Code, Claude, ChatGPT and OpenCode, dictation is written as a prompt. In those apps, terminals and code editors, a selection is only rewritten by Edit Selection, never because you said "fix" or "make".
Busy Electron apps and Chromium browsers get longer to read a paste before your clipboard is put back.
Fixed
Dictation lands where your cursor is in every kind of app: T3 Code, Claude, Cursor, VS Code, Slack and other Electron apps, Chrome and other browsers, and Terminal, iTerm2, Ghostty and Warp, including Claude Code. It goes in the middle of a word or over a selected word too. 1.3.1 could say the text was inserted when one of these apps had ignored it.
After Review before insert or an app switch, ZenVoice waits for the app to be ready before pasting.
Nothing is pasted when you dictate over silence, typing, music or street noise. Engines could write "*music*", "[no audio]" or "The." and ZenVoice pasted it; now the bar says "No speech detected." Words people say alone, such as "Yes." or "Thank you.", still go in.
Nemotron 3.5 no longer loses a whole dictation when it returns an empty word between real ones. About 3 in 100 dictations were affected.
Correcting a time as you say it keeps only the new time: "Meet at 8:30, sorry, I meant 9:30" and "eight forty, I meant nine thirty" give "Meet at 9:30" and "Meet at nine thirty", including "half eight, sorry, half nine" and corrections after a full stop. Before, the sentence could be left unchanged or keep half of the old time.
Times and numbers are no longer offered as Dictionary words, and changing how a time is written in Review does not teach a reusable spelling fix. A number or time already in your Dictionary is marked as not used, so you can remove it; it no longer guides recognition or repair.
No scrollbar appears down the right side of ZenVoice's pages, even when macOS is set to always show scroll bars.
1.3.1
26 October 2026
Dictionary word checking, Kanso formatting, a refreshed app and a more capable command line.
New
More control from Terminal: JSON and streaming transcripts, zenvoice doctor diagnostics, explicit model selection, batch summaries and resumable jobs.
Dictionary corrections for recordings from Terminal, using your saved words read-only or an exported Dictionary file. Background CLI jobs yield while you dictate in the updated app.
Subtitle timing labels and optional experimental English word alignment. Resume checkpoints stay on your Mac as unencrypted, owner-only files.
Parakeet hears your words. With a Parakeet voice model, ZenVoice can now put the words from your dictionary back where Parakeet misheard them, such as "Cloudflur" for Cloudflare or "super base" for Supabase. It never replaces an ordinary word, and works when you dictate in English. Turn it on in Settings → Dictionary.
Listen again for names. An optional 0.16 GB download under Parakeet hears your words checks the recording where a name from your dictionary might have been misheard, such as "Gink" for Nginx. It runs on this Mac, only when a word needs checking, and leaves ordinary English words alone.
See what Parakeet hears your words changed. In History, each word it put back shows what Parakeet heard, such as "Supabase · heard “Superbase”". Take one back and History reads as heard, and ZenVoice keeps that word as heard from then on.
Kanso Satori 1.0, the largest formatting model. Trained on the same examples as Kanso Koi 1.0, it is the most accurate on its own, most of all on dictations unlike those examples; in everyday dictation it writes much the same text as Koi. It is slower than Koi and keeps about 2.7 GB in memory while it works. Choose it in Models.
Better
A darker sidebar with a coloured icon for each page and a clear selection.
About replaces Updates, with version details, update controls, diagnostics and feedback. Stable releases use one update channel.
Help is now at getzenvoice.com/help, with search and links to each answer. Open it from About → Help.
A calmer, flat look on every page. Each group has its title above one plain box, explanations sit behind ⓘ, and pickers show just their value.
In Models, each model shows its Accuracy and Speed bars and its size in the list. Click a model to read what it is good at. Under the active engine, ZenVoice says which voice model it recommends for your language: Parakeet Unified for English, Parakeet Ultra for the other European languages it hears.
In Models, the two sections are now called Voice models and Formatting models.
Models shows who made each model with their own logo: NVIDIA, OpenAI, Qwen, Mistral, Meta and Cohere.
Whisper Small is now a 0.26 GB download instead of 0.49 GB, with the same accuracy. If you already have the larger file, it keeps working until you choose another model.
Whisper Tiny and Whisper Tiny (English) are retired and no longer offered in Models. If you already use one, it keeps working until you choose another model.
ZenVoice's own formatting models are called Kanso. Existing installed models keep working and stay selected under their Kanso names.
Kanso Koi 1.0 is the recommended formatting model in Models and in setup. It is a newer Koi on the lighter base it was trained on: a 1.05 GB download instead of 1.55 GB, about 550 MB less memory, faster formatting and fewer mistakes. Koi 0.9 stays in Models for anyone who prefers it.
Fixed
Word-check download errors appear beside the download control, with retry and cancellation handled there.
Polished keeps the formatting model's cleanup far more often. The check that stops a model from changing your words also threw away correct answers that dropped "like", "you know" or "yeah", or a repeated phrase, without commas around it: about one dictation in five. It now accepts those, and still rejects any other change, including a hedge on a number such as "like 2,000".
1.3.0
22 October 2026
Four new models, more than 1,600 languages, and ZenVoice in your language.
New
Parakeet Ultra. Parakeet V3 trained further: the same 25 languages and the same speed, with fewer mistakes. Your words appear while you speak, and it is now the model ZenVoice recommends for English and 24 other European languages.
Voxtral Realtime. Mistral's model writes while you speak, punctuation included, in 13 languages: English, French, Spanish, German, Russian, Chinese, Japanese, Italian, Portuguese, Dutch, Arabic, Hindi and Korean.
More than 1,600 languages. Meta's Omnilingual ASR hears languages no other model in ZenVoice does, such as Tamil, Bengali, Urdu, Marathi and Swahili. It writes plain words, and ZenVoice's formatting adds the punctuation.
Parakeet Unified. NVIDIA's newest English model, trained for live words and whole recordings at once.
ZenVoice in your language. The whole app, from setup to Settings, in German, French, Spanish, Italian, Swedish, Dutch, Brazilian Portuguese, Japanese, Chinese and Korean. It follows the language of your Mac.
Corrections and lists in ten languages. Say that you meant another day and ZenVoice keeps the one you meant. Count items out loud ("first… second…", "number one… number two…", "step one… step two…") and they become a numbered list, in notes and in Claude, ChatGPT and other AI chats. Only counting does it, so the rest of a prompt stays as you said it.
One tap locks hands-free. With Right Command, Right Option or Right Control, tap once to lock hands-free and tap again to finish, the same as a custom shortcut. Holding the key works as before.
Better
Setup recommends Omnilingual ASR for a language no other model covers, where it used to fall back to Apple's built-in engine. Your current model stays as it is until you choose another.
Setup starts in your Mac's language when ZenVoice can dictate it, and every other language is one menu away.
Transcribe with ZenVoice is switched on in Finder. Before, macOS kept it off until you found it in System Settings.
Try it, the last step of setup, reads its examples in the language you dictate in.
Apple's speech engine uses your Mac's variant of a language, such as Brazilian Portuguese or British English.
Fn works the same way as the right-hand keys when macOS does nothing with the 🌐 key. While macOS uses it for emoji or input sources, a double-tap still locks.
Languages you speak, on the Keyboard page, counts every installed engine, so a Mac with Parakeet alone can add a second language.
Shortcuts are named by the keys on your keyboard. A German keyboard shows the Y key as Y, where it used to show the US letter Z.
Fixed
Double-tap did not lock hands-free on a Mac whose microphone is slow to start. It does now.
ZenVoice deleted some real words it took for fillers, such as German "um" and Spanish "este". It keeps them now.
A double-tap from habit no longer ends the recording its first tap locked.
Asking to translate a selection, or to fix its grammar, no longer fails with "Couldn't make that edit".
On keyboard layouts that put V somewhere else, such as Dvorak and Turkish F, dictated text was not pasted. ZenVoice now follows your layout, for paste and for copying a selection to edit.
1.2.2
15 October 2026
Transcribe recordings on your Mac, and dictate from your mouse.
New
Transcribe a recording. The Transcribe tab turns an MP3, M4A, WAV, AIFF, FLAC, MP4 or MOV file into text on your Mac, with the text appearing as it goes. Dictation always goes first.
Start from the Transcribe tab, drop a file on the menu bar icon, right-click it in Finder and choose Transcribe with ZenVoice, or use Open With. ZenVoice never becomes your default app for any file type.
Copy a transcript or find it saved as a .txt file, switch between raw and formatted text, and export Markdown, SRT or VTT. Recordings have their own model, language and formatting choices.
A zenvoice command for Terminal: install it in Settings → Transcribe, then run zenvoice transcribe meeting.m4a.
Dictate from your mouse. Pick the middle or a side button in Settings → Keyboard: hold to talk, or press once to lock hands-free.
Right Command, Right Option and Right Control can start dictation. Hold to talk, double-tap to lock. Right ⌘ + C and other shortcuts cancel quietly.
A Next Language shortcut, ⌃⌥L by default, moves between the languages you list.
Your last dictation under the notch, off by default. Rest the pointer on the notch to see it, with Copy and Paste again.
Highlight cards for all your stats or one, in three layouts, 12 colours and Wide, Square or Story sizes.
Better
Fn now locks dictation with a double tap, like the other activation keys, so an accidental tap no longer leaves the microphone on. Holding to talk works as before.
A new Keyboard page in Settings holds the activation key, mouse button, shortcuts and languages.
Update feeds must carry ZenVoice's signature, and each update is verified before it's unpacked.
Fixed
ZenVoice no longer pastes into a password field, even when the app hasn't turned on secure input. The text is copied instead.
A password manager's copy on the clipboard is never kept or put back. ZenVoice clears it instead, so the password manager's own clearing still works.
Exporting from Audio History no longer freezes the app.
1.2.1
13 October 2026
Choose your HUD during setup, and dictation that survives app switches.
New
Setup has a HUD step before Try it: pick the notch or the floating panel, and choose whether ZenVoice stays on screen between dictations.
Better
New installs show the HUD only while dictating. Turn on Show ZenVoice at all times in setup, Settings → Dictation or the menu bar to keep it visible. Existing setups keep their current behavior.
Fixed
Switching apps while dictating keeps recording and transcription running. ZenVoice returns to the app where dictation started before inserting the result.
Holding Fn/Globe to dictate is more reliable, including on keyboards that report the key differently. Function and arrow keys no longer start or end a hold.
Changing a shortcut records the keys you press next, including the current shortcut, and ignores a modifier that was already held.
Setup and Settings open at the same window size, and setup steps fit without cut-off content.
1.2.0
11 October 2026
Softer speech, safer formatting and more reliable dictation.
New
Quiet speech mode in Settings → Audio is tuned for softer input. A microphone check compares room noise with your speaking level without saving a recording.
Fixed
More reliable dictation insertion in native text fields, including TextEdit.
Canceling a dictation discards late results and obsolete previews. A new dictation starts with a clear preview.
Formatting keeps meaningful qualifiers, names, numbers, dates and negations. Spoken corrections and voice edits reject changes that alter your facts.
History reset, retention and recovery handle overlapping work more reliably. Insights distinguishes recorded activity from incomplete or missing metadata.
Model preparation and formatting share GPU resources safely when requests are canceled or time out.
Manual update checks work while automatic checks are disabled, and Settings explains how installers are verified.
1.0.9
9 October 2026
A clearer setup and a dictation test that works.
Better
Onboarding now matches the app’s headers, cards and typography, with calmer animation and readable model labels.
Download controls show progress, verification, preparation and waiting status, with clear cancellation and retry actions.
Fixed
Setup waits for the speech engine you chose to finish downloading and preparing. Formatting downloads wait for other downloads instead of silently stopping.
The dictation test has Start/Stop controls, microphone activity and inline errors. Your words land directly in its text box, including with copy-only or review-before-insert enabled. Leaving the test discards late results.
Shortcut recording stops when its screen closes. Fresh installs can choose Hinglish or automatic detection before downloading a model, with engines suited to the chosen language.
Apple Speech checks authorization and prepares the selected language’s system assets before setup reports it as ready.
1.0.8
5 October 2026
Model downloads recognize their approved sources.
Fixed
Approved Parakeet, Nemotron, Cohere, Qwen and formatting model downloads no longer fail with an unapproved-source error. Pinned revisions, file sizes and SHA-256 checks are still required before installation.
Kanso’s nested adapter files download and install correctly.
1.0.7
2 October 2026
Your real microphone during calls, and cleaner commas.
Fixed
When Zoom, Teams, Krisp or an audio router makes its virtual input the macOS default, ZenVoice now records from your real microphone instead of getting silence or call audio: the built-in microphone first, then a USB or Bluetooth one. A microphone you pick in Settings → Audio is always used as chosen, and Settings says when the swap happens.
No stray comma where an “um” or “uh” was removed: “and, uh, anxiety” is now “and anxiety”. A comma that marks a real pause stays: “in two weeks, repeat the blood work”.
1.0.6
30 September 2026
English that arrives sooner and reads cleaner, with nothing heavy in memory.
Better
English is cleaned up by ZenVoice Rules by default: your text appears sooner after you let go, and no formatting model sits in memory.
More fillers come out: “basically” anywhere, “you know”, openers like “yeah so”, and a filler “like” (“we need to figure out like how to bring sales” → “We need to figure out how to bring sales.”). A real “like” stays: “I would like if it was somewhere warm”, “things like formatting”, “Like I said”.
The rules never add or reword a word, so names, numbers and commands come through exactly as you said them.
“ZenVoice Rules” is the new choice in Formatting → Formatting model and the recommended one in setup. Kanso, Qwen 3.5 and Apple Intelligence are still there, and if you chose Kanso it stays selected.
1.0.5
29 September 2026
156 emoji you can say out loud, and voice edits that make proper lists and never change your facts.
New
115 more spoken emoji, 156 in all: say the name and “emoji”, as in “crying laughing emoji” 😂, “heart eyes emoji” 😍, “handshake emoji” 🤝 or “India flag emoji” 🇮🇳. Faces, hands, hearts, food, travel, weather, animals and flags.
Hinglish emoji names: “dil wala emoji” ❤️, “rone wala emoji” 😭, “tiranga emoji” 🇮🇳 and 34 more. An emoji only appears when you say “emoji”, so “I was crying laughing” stays as words.
Better
Edit Selection makes proper lists: “bullet points” and “numbered list” put each item of a sentence like “update the page, fix the bug, and email the users” on its own line, keep a heading such as “We need to:”, and never merge or reword items. They're instant.
“Make it shorter” cuts more, while keeping every fact.
Voice edits never change your facts: an edit goes in only if every number, name and “not” is still there, and Hinglish stays Hinglish. Otherwise your selection is left as it was.
When there's nothing to fix, ZenVoice says so (“Nothing to change”, “It's already short”) instead of “Couldn't make that edit”.
1.0.4
29 September 2026
Hinglish that arrives in seconds, reads cleanly and says only what you said.
Better
Long Hinglish dictations arrive much sooner: Hinglish Prime transcribes while you talk, so letting go of the key waits only for the last few seconds, even after a couple of minutes of speech.
Hinglish is cleaned up by rules made for it: numbers as digits (“saadhe chaar baje” → 4:30 baje, “pachchis hazaar rupe” → ₹25,000, “das percent” → 10%), question marks, and a comma after “Haan ji” or “Are yaar”. Kanso, trained on English, no longer runs on Hinglish.
Fixed
Short Hinglish dictations no longer start with made-up words (“Theek.”, “The.”) or slip into English.
Hindi repetition is kept: “kabhi kabhi”, “baar baar” and “bahut bahut dhanyavaad” were being cut to one word.
Long Hinglish dictations no longer run two sentences together.
1.0.3
28 September 2026
Edit any selected text by voice, long dictations that arrive in a couple of seconds, and a Memory setting.
New
Edit Selection: select text in any app, press ⌃⌥E, say the change (“make it shorter”, “bullet points”, “fix the grammar”) and press again. Change the shortcut in Dictation → Shortcuts.
Models → Memory: choose how long models stay loaded after you dictate. Lowest memory frees them as soon as your text is in, Balanced (the default) after 30 seconds, Fastest after 5 minutes.
Better
Long dictations arrive much sooner with NVIDIA Parakeet: speech is transcribed while you talk and Kanso formats a few sentences at a time as you go, so letting go of the key waits only for the last part. Long text is also formatted more accurately.
The Professional tone no longer holds up long dictations (over 300 words): they get the same text straight away.
When Kanso runs out of time, your text goes in at once, and its memory no longer grows over a long session.
Fixed
Voice edits now work in browsers and Electron apps, and on short instructions such as “shorter” or “bullet points”, instead of typing the words after your selection.
1.0.2
27 September 2026
Translate to English now works with every speech engine.
Better
Translate to English prefers an installed Whisper model that translates on its own (Large V3, Medium, Small, Base or Tiny), fully offline. Otherwise Apple Intelligence translates on your Mac.
The Translate to English setting now says how it will translate on your Mac, or what's missing.
Fixed
Whisper Large V3 Turbo and the Hinglish models write the language you speak, so with Translate to English on their text is now translated afterwards instead of being pasted as spoken.
1.0.1
27 September 2026
Your first dictation works during setup, and Pro keys can be upgraded from the app.
New
Upgrade to Lifetime… on the License screen for Pro keys.
A quiet licence check about once a week while a licence is active, so refunded keys stop working. Offline, nothing changes.
Fixed
Setup's last step, “Try it”, now starts your 14-day free trial, so your very first dictation works.
1.0.0
27 September 2026
Launch
Private dictation for Mac: speak, and clean text lands in any app, transcribed and polished on your Mac.
What's included
Press ⌃⌥Space, or hold the Globe (Fn) key, and talk: your words are typed into whatever app you're in. ⌃⌥V pastes the last dictation again.
Eleven on-device speech engines: NVIDIA Parakeet TDT 0.6B V3 and V2, NVIDIA Nemotron 3.5 Multilingual, Whisper Large V3 Turbo, Small and Tiny, Cohere Transcribe, Qwen3-ASR 0.6B, Apple Speech, and Hinglish Prime for Hindi and English written as Hinglish.
Kanso Koi 0.9, ZenVoice's formatting model: fillers and false starts out, punctuation and capitalization fixed, numbers, money and dates written the way you'd type them. On 500 held-out dictations it has a fifth of the WER of the stock model it's trained from.
Apple Intelligence and Qwen 3.5 2B for Smart formatting too.
Translate to English: speak another language and get English, on your Mac.
The notch HUD, a live capsule that grows out of the camera housing, and the pill, a compact HUD that works on any display.
Insert, copy only, or review before insert, and your clipboard restored after pasting.
Voice snippets: say a trigger phrase and ZenVoice types the full text.
Encrypted history with search, retry and a Recovery Inbox for dictations you cancelled.
Insights: words per minute, app rankings, recording time and a streak calendar, all computed on your Mac.
A voice profile with your own correction rules, Audio Doctor, and a Permissions screen that shows exactly what ZenVoice can access.
Model downloads pinned to a reviewed revision and checked by SHA-256 before they load.
A 14-day free trial, then buy once: Pro, Lifetime or Teams. Activation needs the internet once; after that ZenVoice works offline.
Automatic updates, signed and checked before they install.
Requires macOS 15 or later. Apple Speech needs macOS 26.