What's new

Give any mode its own hotkey. The Modes tab now gathers every mode — the built-ins and your own custom ones — in one place, and you can assign an optional hotkey to each. Press it and that single dictation runs in that mode (say, Professional Email), then you're right back to normal: your everyday hotkey and your active mode don't change. None sits at the top and owns your main dictation hotkey.

Dictation stays fast no matter how much history you've built up. If recording had started to feel sluggish after weeks of heavy use, this is why — and it's fixed. Every dictation used to re-save your whole history to disk dozens of times a second, so the more history you'd accumulated, the slower it got. Recording speed no longer depends on how much history you've kept. (History now holds your most recent 500 dictations.)

The floating pill is tidier. When you stop talking, the pill now shrinks back to a neat round orb while it transcribes — it used to stay stretched out with an empty gap where the timer had been — and the processing orb sits properly centered.

Meeting Notes mode now has a fixed, predictable layout. A short summary, a flat "Key points:" list, and — only when you actually mentioned any — a separate "Action items:" section. Previously the AI was free to improvise the structure, so bullets sometimes nested inside other bullets and action items occasionally hid inside the discussion points. If you use HyperVoice Cloud for AI clean-up, you get this fix immediately without updating; with your own OpenAI/Anthropic key, it ships in this version.

Live typing now streams in Chinese and Japanese. Languages written without spaces used to wait for the entire sentence before anything appeared — now the text commits character by character as you speak, the same way English commits word by word. (And no stray spaces sneak in where those languages don't use them.)

Word counts are now correct in those languages too. A Chinese or Japanese dictation used to register as roughly one "word" no matter how long it was, which made your statistics meaningless — and quietly meant the free tier's daily limit never applied. Each character now counts the way you'd expect, so your stats are real (and the daily limit works the same for everyone).

Less ghost text when you pause. The speech engine sometimes invents a stock phrase out of silence — in English we already filtered the classics like "thanks for watching," and now we filter the well-known equivalents in Spanish, French, German, Japanese, Chinese, Korean, Arabic, and half a dozen more.

A heads-up for Chinese / Japanese / Korean keyboards. The default hotkey (Ctrl+Shift+Space) can collide with your input method's language-switch shortcut on Windows. The hotkey settings now warn you when that's the case and suggest a combo that won't fight your IME.

No more "the hotkey just died." On the free tier, hitting the 500-words-a-day limit used to silently ignore your hotkey — it looked broken. Now you get a Windows notification the moment a press is blocked, and the floating pill shows "Daily limit — upgrade or wait."

Account linking always offers a way out. If the browser hand-off fails for any reason, the linking screen now always shows a copyable sign-in link you can paste into any browser — previously it only appeared for some errors.

The website caught up to the app.

The orb is the hero now. The landing page opens on the same living voice orb from the app — and it's real: click it, allow your mic, and watch it react to your own voice, right in the browser. Nothing is recorded and nothing leaves your browser; it's the app's animation, reacting live. (If the mic is blocked it performs a simulated take instead.)

Walk through the Showroom. A new "Enter the showroom" button takes you into a first-person, walkable 3D gallery of the whole HyperVoice suite — WASD and mouse, museum spotlights, a polished reflective floor, framed brand art, story placards, fun facts, and a release-timeline wall. The big screen in the HyperVoice wing runs the live orb.

How It Works tells the story as you scroll. Press the key (it presses), speak (the waveform reacts), and your words land at the cursor — with a pixel-accurate preview of the floating pill exactly as it behaves in the app today.

Honest words, lighter page. Every section was rewritten in plain first-person language, the announcement bar calmed down, and the privacy wording is precise everywhere: the speech engine runs on your device; the optional AI clean-up step is the part that uses the cloud.

The app behaves at its smallest size. At the minimum window size, the Record screen no longer clips the orb out of reach when the result panel is open, and the metadata chips in History and on the Record screen wrap neatly instead of squeezing the buttons off the edge.

One voice everywhere. A full pass over the app's wording: they're "dictations" consistently now, headings and buttons share one style, and ellipses are real ellipses. Dozens of tiny mismatches, all aligned.

You can always tell when the mic is live. The system-tray icon now shows a red recording dot while HyperVoice is listening (and the tooltip says so) — a glanceable safety check even when every window is hidden. If the mic fails mid-recording, the indicator clears rather than sticking.

Optional sound cues. A new toggle in Settings → Dictation plays soft ticks on record start and stop and a gentle blip when your text lands — handy when dictating into a window you're not looking at. Off by default, so nothing changes unless you want it.

Keyboard and polish. Every button in the app now shows a clear focus ring when you Tab to it; the title bar controls feel native (including a proper red close-button hover); and Statistics celebrates milestones — cross 1k, 10k, 50k, 100k words and the page says so, with progress toward the next badge.

Empty dictations no longer clutter History. If a recording came back with nothing (silence, no words), it no longer leaves a blank card in History — the explanation banner tells you what happened instead.

A fast follow to the redesign, focused on how the app feels.

Lighter when you're not looking. The animated orb and waveforms now pause completely while their window is hidden and run at half rate while idle — so HyperVoice sits closer to zero in Task Manager when it's just waiting for you.

A little celebration. When your dictation finishes, the orb now fires a quick green triple-ripple — a half-second "done" you can catch from the corner of your eye.

Live typing (experimental) now shows its preview in the pill. If you've turned on live typing, the floating pill streams the words it's hearing as you speak — solid for the settled part, faded for the part still being worked out.

Reliability, continued:

- Changing the speech engine's settings mid-dictation can no longer quietly bring back the old model on later dictations (now covered for Parakeet as well as Whisper).

- Dictating while the speech engine is still warming up no longer risks two engine loads colliding on the GPU — the dictation now waits for the warm-up and uses its result.

- A GPU crash during the onboarding mic test is now caught the same way as during real dictations, so the automatic switch-to-CPU protection kicks in from your very first minute with the app.

The whole app caught up to the orb. This is the largest visual release HyperVoice has had — every screen, one design language.

One living design language. The voice orb's colours now run through the entire app: cool blue when idle, hot red-orange while listening, amber while transcribing, violet while polishing. The Record screen's background breathes with it, charts and chips pick it up, and every card in the app shares one "glass" material with proper depth. Even the scrollbars stopped looking like a website.

Settings, reorganised. Six tabs that match what you're actually looking for: General, Dictation (hotkey, mic, recording mode, the try-it box), Speech & Models (engine, models, GPU, language — no longer buried), AI Cleanup (providers and keys, with styles managed in one place: the Modes page), Pill, and Account & About. Everything works exactly as before — it's just where you'd expect it now.

History and Statistics, rebuilt. History entries are proper cards with timestamps ("2h ago"), word/speed/duration chips, and a search that breathes. Statistics now leads with the number that matters: time saved versus typing, from your own words and speaking speed — plus a day-streak counter and your best day on record.

Onboarding got the cinema treatment. The mic test sits on a dark stage that glows with your voice, progress slides instead of blinking, and the model download finally looks like something worth waiting for.

Light theme, actually finished. A full audit fixed colours that washed out or disappeared on light backgrounds — including a long-standing bug that quietly disabled every dark/light style variant in the app.

Plus the polish: the sidebar's active marker slides between tabs, screens fade in instead of snapping, diagnostic messages float over the top (and auto-dismiss after 10 seconds) instead of shoving the layout down, the orb holds perfectly still when you start recording, and contrast got a real pass — no more dark-grey-on-dark-grey.

A fast follow-up to this morning's release, fixing the most-reported pill annoyance.

The pill no longer blanks out when your mouse gets near it. Since v0.5.164, the pill has let clicks in the empty space around it pass through to whatever's behind — useful, but the way it flipped that switch could make Windows' renderer blank the pill entirely: the window was still "there", just drawing nothing, usually right as your cursor approached. It now flips only the single setting that controls mouse pass-through and leaves the rendering path alone, so the pill stays drawn. Two smaller guards landed with it: the pill now recovers automatically if a resolution change or monitor unplug pushes it somewhere off-screen, and it reclaims its always-on-top spot even from apps (like streaming overlays) that pin themselves above it.

The biggest visual update HyperVoice has had, plus a deep pass on dictation reliability.

The Record screen has a new heart. The flat circle is gone. In its place: a living plasma orb that reacts to your voice in real time — a spectrum ring that spikes as you speak, particles that orbit faster the louder you get, and pulses that ripple outward on every emphasis. It changes character with what the app is doing: cool blue while waiting, hot red-orange while listening, an amber sweep while transcribing, violet while polishing with AI. Same big button, same hotkey — it just looks alive now.

The floating pill got the same soul, minus the clutter. While you dictate, the pill is now minimal: a miniature version of that orb on the left, the timer on the right, nothing else. The pill's outline also shifts colour with the state, so you can tell what it's doing from the corner of your eye. And you can now pick what the idle pill looks like in Settings → General → Idle Pill Style: the classic logo + HyperVoice name, or just a tiny breathing orb — the smallest the pill has ever been.

The pill no longer vanishes. If you had it set to "always visible", the pill could disappear after pressing Win+D, switching out of a fullscreen app, or unplugging a monitor — and stay gone until your next dictation. It now notices within a few seconds and brings itself back (without stealing your focus). Saved pill positions also survive multi-monitor setups with different display scales, and a position saved on a since-unplugged monitor no longer strands the pill off-screen.

Dictation is a lot harder to break. A reliability pass over the whole pipeline:

- Accidentally double-pressing the hotkey can no longer paste your dictation twice or leave a recording running that nothing could stop.

- An engine error mid-transcription can no longer leave the app stuck on "Transcribing" with a dead hotkey — every failure now resets cleanly and tells you what happened.

- If the window you were dictating into closes before your words are ready, HyperVoice no longer types them into whatever window happens to be in front. Your text goes to the clipboard instead, with a note to press Ctrl+V where you want it.

- If your microphone disconnects mid-recording, you now get told immediately instead of the recording silently going dead.

- AI cleanup with your own OpenAI/Anthropic key can no longer hang forever on a bad connection — it times out and your raw transcription is kept.

- Apps that receive dictation as typed keystrokes now get real Enter presses for line breaks, so multi-line results (like the Email style) format properly.

Experimental: live typing. There's a new opt-in toggle in Settings that types your words into the target as you speak, then settles the final wording when you stop. It's early and off by default — if you try it, tell us how it goes.

A round of security hardening around how the desktop app links to your HyperVoice account.

The app now asks before linking from an unexpected web link. Signing in works by having your browser hand a hypervoice:// link back to the app to link this device. The app used to act on that link no matter where it came from. Now, if such a link arrives without you having just started a sign-in yourself, HyperVoice asks you to confirm first — so a random web page can't quietly re-point your app at someone else's account. Signing in the normal way (click Sign In, then confirm in your browser) links with no extra prompt, exactly as before.

Your signed-in session is now tied to this device. Behind the scenes, your account session can no longer be reused from a different machine if it ever leaked. Nothing to set up — it just works.

Two things in this one.

Dictation now lands in apps that used to ignore it. A few apps — custom terminals, and some built on web tech (Electron / Tauri-style webviews) — read the clipboard in a way that raced how HyperVoice pastes: by the time the app looked for your dictation, HyperVoice had already restored your previous clipboard, so nothing appeared. HyperVoice now spots these apps and types your words straight into them instead of pasting, so they show up reliably. Ordinary apps and your browser keep using the fast paste they always did. Nothing to set up — it picks the right method for each app on its own.

A simpler Record screen. The Record tab is now built around one clear action: a big mic button, your hotkey shown as actual keys ("press Ctrl Shift Space and talk"), and a reminder that your words land wherever your cursor is. Your processing styles moved into a single Style button — click it for the full picker (Clean Up, Email, Notes, and any custom styles you've made) — instead of a wall of cards. Less to take in, all the same power.

The floating pill sits inside a slightly larger invisible window so its glow has room to fade out. The catch: that invisible margin around the pill was quietly catching your clicks — so if you clicked a button or link sitting within ~20px of the pill, nothing happened, and you'd have to drag the pill out of the way first.

Now only the pill itself catches the mouse. Click anywhere in the empty space around it and the click goes straight through to whatever's behind it — while dragging the pill to move it, and right-clicking it for the menu, both still work exactly as before.

A few people noticed dictation getting much slower out of nowhere. Here's what was happening, and the fix.

HyperVoice transcribes on your GPU — roughly 5× faster than the CPU fallback. As a safety net, if the app ever crashed mid-transcription it would switch the GPU off on the next launch and run on CPU instead, assuming your graphics drivers were to blame. The problem: it only took one crash to switch the GPU off, it stayed off permanently, and it happened silently — so if a single odd recording ever tripped it, every dictation afterwards quietly ran on the slow CPU path and you'd have no way to know why.

Fixed three ways:

Your GPU comes back on its own. If a past crash had left you stuck on CPU, HyperVoice re-enables GPU acceleration the next time it launches and shows a note telling you so — dictation should feel fast again straight away. (Prefer CPU? You can still turn the GPU off any time in Settings → Dictation.)

One hiccup no longer benches your GPU. A single crash is now treated as the fluke it usually is and your GPU stays on. HyperVoice only falls back to CPU after several crashes in a row with no successful dictation in between — the real signature of a genuinely broken GPU, not a one-off.

A stuck recording can't run away. The thing that started all this was a recording that somehow ran for over an hour — usually a hotkey that got stuck, or a toggle recording left running. Recordings longer than 30 minutes are now stopped with a clear message instead of being handed to the engine, which is what crashed it in the first place.

Nothing to do on your end — install the update, and if you'd been affected you'll see the "GPU re-enabled" note on first launch.

Two changes to the floating pill:

New — keep the pill out of the way until you need it. Settings → General → Pill Visibility has a new option: switch from "Always visible" to "Only while dictating" and the pill stays hidden until you start a recording. It appears the moment you press your hotkey, stays up while it's transcribing/processing, shows the ✓ when it's done, then disappears again. It also still pops up for alerts (like "this app blocked the paste — press Ctrl+V"). When it appears mid-dictation it won't steal focus from whatever you're typing into. Leave it on "Always visible" (the default) and nothing changes.

Fix — black rectangle behind the pill. On some setups a solid black box was showing behind the rounded pill instead of letting your desktop show through. The pill is properly transparent again.

Building on v0.5.160: the installer now closes any running copy of HyperVoice itself, before it installs or uninstalls. So "Update Now" and Windows uninstall both work cleanly every time — no more "Error opening file for writing", and no leftover that won't delete. Because the installer handles this, updating onto this version works even from older builds.

Updating from inside the app could fail with a Windows "Error opening file for writing" message and never finish — the app stayed running in the tray, so the installer couldn't replace it. HyperVoice now fully closes itself for the updater, so "Update Now" installs cleanly.

One-time step: because the old behaviour is part of the version you're on right now, this particular update needs to be installed manually once — download HyperVoice from hypervoice.app with the app fully closed, then run it. After you're on this version, future updates install automatically from the in-app button.

If you dictate into an app running as administrator, Windows won't let HyperVoice auto-paste into it (it blocks keystrokes sent from a normal app into an elevated one) — so the text stays on your clipboard and the pill tells you to press Ctrl+V yourself. That message now displays cleanly: it's centred in the pill and no longer overlaps the "HyperVoice" wordmark.

v0.5.152 already switches you to a multilingual model when you pick a non-English language (or Auto) while on an English-only model. This adds a second safety net at the moment you actually dictate: if anything slipped through — for example, you downloaded a multilingual model after picking your language — HyperVoice now catches it right before transcribing.

If a downloaded multilingual model is available, it switches to it automatically. If not, instead of quietly transcribing your French (or other non-English) speech as English gibberish, it stops and tells you to download a multilingual model in Settings → Models. English-only dictation is completely unaffected.

The v0.5.155 fix stopped the speech engine from flipping your language to English. But there was a second place the same thing could happen: if you ran a HyperVoice Cloud cleanup mode (Clean Up, Professional Email, Bullet Points, and so on) on something you dictated in French — or German, Spanish, Japanese, any non-English language — the cleanup step itself could quietly rewrite your text into English.

The reason was that the cleanup instructions were written in English, and on a short phrase the AI would "tidy up" by translating rather than just cleaning. Now the cleanup step is explicitly told to answer in the language you spoke and not to translate — so French stays French, Spanish stays Spanish, and so on. Auto-detect follows whatever language it detected.

If you bring your own OpenAI or Anthropic key, you were already covered by the v0.5.155 fix — this closes the same gap for HyperVoice Cloud. English dictation is unchanged, and there's nothing to update: it's a server-side fix that's already live.

If you're on Ubuntu, Fedora, Mint, Pop!_OS, or any other recent x86_64 Linux desktop, you can install HyperVoice today. Download a single .AppImage file from the dashboard, make it executable, double-click, and the same speak-with-a-hotkey experience Windows users have lands at your cursor in whatever app you're typing into.

How to install. Grab HyperVoice_0.5.157_amd64.AppImage from your dashboard (or directly from hypervoice.app/api/download/linux). Right-click → Properties → Permissions → "Allow executing as program", or chmod +x it from a terminal. Then double-click to launch. The first run takes you through model download + hotkey setup, same as Windows.

Requires an X11 session. Wayland users — pick "Ubuntu on Xorg" (or "Plasma X11", or your distro's equivalent) at the login screen for now. HyperVoice needs to restore focus to the window you were typing in before injecting Ctrl+V, and Wayland deliberately blocks cross-app focus changes without a desktop-portal prompt. That prompt would fire on every dictation and ruin the experience, so we ship X11-only until the portal situation improves. There's a walkthrough on the install help page covering this in more detail.

Whisper engine only, for now. The optional Parakeet engine (Settings → Dictation) is hidden on Linux because it depends on DirectML, a Windows-only graphics API. Whisper handles all the same languages and runs on Vulkan GPU just like the Windows build — most users won't notice the difference.

Beta-quality, deliberately. The Linux build has been compile-tested, UI-tested, and install-tested, but it hasn't seen wide real-world use yet. Audio capture goes through ALSA/PipeWire which is the standard Rust audio path and should "just work," but if you hit a snag — wrong mic picked up, GPU not detected, paste landing in the wrong window — drop a note in Discord or email support@hypervoice.app and we'll patch it fast. The "beta" tag comes off once a few of you have put it through its paces.

Dashboard polish for everyone (Windows + Linux):

- The Download button in the dashboard tab bar is now a split button — click the caret to install on a different OS (Windows ↔ Linux).

- The "Get Started" checklist can be dismissed with an × — for returning users who don't need the hand-holding. Bring it back from Account → Dashboard if you change your mind.

- The Discord + What's New pills now sit on one tidy row at the top of the dashboard instead of two stacked blocks.

Three small improvements to the recording pill — the little floating bar that appears when you press the hotkey.

You'll see "HyperVoice" on it now. Next to the logo there's a "Hyper" + "Voice" wordmark (cyan + orange, matching the app icon). Two reasons: it makes the pill recognisable as HyperVoice at a glance, and it gives it light-coloured pixels that stay visible on any wallpaper — including black desktops, where the previous all-dark pill could blend into the background. There's also a subtle orange ring around the pill now so it has a visible edge regardless of what's behind it.

The pill stays on top, even after fullscreen apps grab focus. Previously, if you played a fullscreen game, ran an installer, or opened an admin-elevated dialog, the pill could end up hidden behind those windows and only reappear the next time you pressed the hotkey. HyperVoice now re-asserts "always on top" every 3 seconds in the background, so within a moment of returning to your desktop the pill is back where it should be.

You can drag the pill flush to the top of your monitor. Used to be that dragging it as high as it would go left a ~24-pixel gap between the pill and the screen edge — felt like the pill was "snapping down." That gap is now ~8 pixels, just enough room for the glow around the pill not to clip at the screen edge.

If you set your Speech Language to French (or German, Spanish, Japanese — anything other than English) and dictated, you might have noticed the text came out as English instead of the language you spoke. Same thing on Auto-detect: even when HyperVoice correctly recognised that you were speaking French, the output text would silently flip to English.

This wasn't translation in the helpful sense — it was a Whisper bug where the model was being given an English-language style hint that overrode the language you'd chosen. The hint is gone now for any language other than English, and your spoken language comes out as written text in that same language, the way it should have all along.

Nothing changes for English-only dictation — the punctuation and capitalisation hints are still in place there.

The Sign out button under Settings → Account used to fully unlink your HyperVoice account from this machine — so signing back in meant going through the browser bounce-back flow again every single time. That's been split into two buttons that do what their names suggest.

Sign out (the primary button) now clears your active session in HyperVoice without removing your credentials from the machine. The next time you launch the app and tap to sign in, you're back in immediately — no browser bounce, no setup code, no waiting. Useful if you just want to step away, hand the keyboard to a co-worker for a meeting, or close out at the end of the day without losing your link.

Switch account (a quieter secondary button next to it) does what Sign Out used to do — fully removes your credentials from this device. Use it when you actually want to sign in as a different person, or when you're handing the machine over to someone else and don't want them inheriting your session.

A heads-up on the trade-off. Because Sign Out no longer wipes credentials, anyone with access to your Windows user account could in principle launch HyperVoice and dictate against your account without re-authenticating. The credentials are still encrypted by Windows at rest (the same protection Slack, Discord, and Spotify rely on for their desktop apps), so it's not a free-for-all — but if you share a computer with other people who can log into your OS user, use Switch account rather than Sign out when you're done.

Nothing changes if you've never signed in, and nothing changes for legacy license-key users — this only affects the unified account sign-in flow.

Open Settings → AI (or click the language in the status bar) and you'll now see a new Auto-detect option pinned at the top of the language list, above a separator. Pick it, and on every recording Whisper figures out what language you spoke and transcribes accordingly. If you've turned on cleanup or any other post-processing mode, it follows the detected language too — so a quick Spanish dictation gets cleaned up in Spanish, a French one in French, without having to switch the language picker each time.

A heads-up on accuracy. Whisper detects the language from the first ~30 seconds of audio. For longer dictations and major languages (Spanish, French, German, Mandarin, Japanese, etc.) it's reliable. For very short clips of 5–10 seconds, or for languages that sound similar to each other (Spanish/Italian, Mandarin/Cantonese), picking the specific language is still more accurate. If you mostly dictate in one language, sticking with that explicit choice gives the best results — Auto-detect is built for the case where you genuinely switch languages.

Requires a multilingual model. The English-only Whisper models (Tiny English, Base English, Small English, Medium English) are trained without the language token, so they can't participate in detection. If your current model is English-only and you pick Auto, HyperVoice automatically switches you to a downloaded multilingual model if one's available, or prompts you to download one if not.

The status bar shortcut shows "Auto" while it's active, and the popover lists Auto-detect at the top regardless of which specific languages you've recently used — so it's always one click away.

One-click download if you don't have a multilingual model yet. Auto-detect needs one of the multilingual models. If you only have an English-only model installed (say, Base English) and you pick Auto-detect or a non-English language, the yellow warning above the model list now includes a primary "Download Base Multilingual (142 MB)" button — the suggested model always matches the size of the English-only model you already have, so you don't get pushed into a 1.5 GB download if you're on Tiny English, and you don't get nudged toward Tiny if you're already happy on Medium English. The button shows progress inline; once the download finishes, the new model becomes active automatically. If you want a different size (like Large v3 Turbo for top accuracy), "or pick another below" sends you straight to the full model list.

Multilingual models are now labelled "Multilingual". Previously the model picker showed "Base English" paired with a bare "Base", which read as "what's the difference?" The multilingual variants are now displayed as "Tiny Multilingual", "Base Multilingual", "Small Multilingual", and "Medium Multilingual" — symmetric with their English-only siblings. The Large models keep their existing names because there's no English-only Large to disambiguate from.

If your speech language is set to English, the model picker in Settings → AI now hides the multilingual Tiny, Base, Small, and Medium models. Those rows were never the right pick for English — for each of them, there's an English-only sibling (Tiny English, Base English, etc.) that's faster and more accurate at the same size. Showing both side-by-side just invited downloading the worse one.

The Large models (Large v2, Large v3, Large v3 Turbo) stay visible on the English list. They're multilingual-only — there's no "English-only Large" because at that size the multilingual model already does English as well as anything would. If you want the highest accuracy tier, that's still where to look.

Switching your speech language to anything other than English continues to work the same way: only the multilingual models are shown, since the English-only variants can't transcribe other languages.

HyperVoice used to silently download and restart itself when a new version was available. That's gone.

You're in charge of when to update. When a new version drops, you'll now see a small banner at the bottom of the app — right above the status bar — that says "Update v0.5.X available" with three buttons: What's New, Update Now, and a close-X. Nothing happens until you click one. Close the app, leave it running for a week, ignore the banner indefinitely — your call.

See exactly what you'd be installing. Click What's New and the changelog opens scoped to entries newer than your installed version — no scrolling past releases you already have. There's a "Show all" link if you want to browse the full history.

Dismiss it for now without it nagging you forever. The X hides the banner for the current available version only. When the next release ships, the banner reappears for that one — so you won't accidentally miss a critical fix months from now just because you closed it once.

If you change your mind and want to update after dismissing, Settings → About still has the manual "Check for Updates" button it always did.

Spanish, French, German, Portuguese, Italian, Japanese — if your computer's regional setting is anything other than English, the first-run setup now shows a language picker before downloading the speech model. Picking your language gets you the right multilingual model from the start, and the mic-test step uses a localized test sentence so you're reading words in your actual language rather than a phonetic-sounding English phrase.

Nothing changes for English-locale users — onboarding stays the same one-click flow as before.

This closes a gap where non-English speakers were silently landing on the English-only model and getting empty transcriptions on their very first dictation. If you've already been through onboarding and you'd rather pick a different language, Settings → AI has the full 99-language picker that shipped in v0.5.148.

If you dictate in anything other than English, this release is for you.

A real language picker in Settings → AI. Type-to-filter through all 99 languages Whisper supports — Polish, Russian, Korean, Arabic, Catalan, Swahili, whatever. Native names show alongside the English label, so you can find your language by typing it the way you spell it locally. The old "stuck on English with no way to change" situation is gone.

Models that don't fit your language get out of the way. If you pick anything other than English, the English-only (.en) speech models disappear from the model list — they literally can't transcribe other languages, so showing them was just noise. If your active model happens to be an English-only one when you switch language, HyperVoice will silently swap to the closest-sized multilingual model you already have downloaded (so a Base English user switching to French lands on Base multilingual if it's on disk, or Small if not). If nothing multilingual is downloaded, you get a clear "download a multilingual model" prompt rather than silent breakage.

Quick-switch from the status bar. A new globe icon at the bottom of the window shows your current language. Click it and you'll see a popover with the 2-3 languages you've actually been using recently — switch between them with one click. "More languages..." in the same popover jumps straight to the full Settings picker if you need it.

Post-processing now stays in your language. If you use Clean Up, Professional Email, Meeting Notes or any of the other AI cleanup modes, they used to occasionally translate short non-English dictations into English — because the instructions sent to the LLM were written in English and the model would helpfully "match the prompt." Now every cleanup mode gets an explicit "the user is dictating in {your language}, respond in {your language}, do not translate" prefix. Whether you're on HyperVoice Cloud (Pro) or your own OpenAI/Anthropic API key, your French stays French, your Polish stays Polish.

Custom modes you've written yourself are left alone — that's your prompt, you own it. If you want them to be language-aware, write that in.

If you'd linked your HyperVoice account on this machine, there's now a Sign out button under Settings → Account → Link Web Account. Use it to switch accounts, sign out before handing your machine to someone else, or wipe your linked state without manually deleting credentials. The hotkey + your local settings keep working after signing out (you'll be on the free tier until you re-link).

If you dictate in a non-English language, the download step during onboarding now tells you up-front that you can switch to a multilingual speech model in Settings → AI anytime — no more guessing whether you're stuck on English.

If you install HyperVoice for the first time, you'll no longer see the "choose your language" step. HyperVoice defaults to English with the Base English speech model and starts downloading it in the background the moment you launch the app — so by the time you've read through the welcome screen and hit Next, the model is usually already on disk and you can dictate immediately.

If you dictate in a non-English language, head to Settings → AI after onboarding and pick a multilingual model. The picker is gone from onboarding, not from the app.

If you visit your dashboard on Android, iOS, macOS, or Linux, the orange Download buttons no longer appear — HyperVoice today is a Windows-only desktop app, so offering you an installer you can't run wasn't useful. Instead you'll see the existing platform-waitlist banner, and we'll email you when a build for your operating system ships.

If you've already clicked the download button on a non-Windows device in the past, the .exe you got won't run — nothing has changed there, sorry. The waitlist already has your email logged from your first dashboard visit.

If you use a processing mode (Clean Up, Email, Bullet Points, etc.) with HyperVoice Cloud as the provider, every result and every history row now shows a small +XXXms cleanup chip next to the existing recording-time chip. Now you can see at a glance how long the cloud LLM step took, separate from the local transcription time.

Raw dictations (no processing mode) look the same as before — the new chip only appears when there was actually a cleanup step to time. Past history entries from before this update also won't show the chip because they didn't record the value; new dictations will, going forward.

If the mic test in onboarding picks up no audio or near-silence, the Next button now changes to a quieter "Continue anyway" and asks for explicit confirmation before letting you move on. The confirm dialog points at the most likely cause (wrong input device selected, hardware mute on a headset, another app holding the mic) and the device picker on the same screen lets you switch mics without leaving onboarding.

Previously it was easy to click past a failing mic test without realising — and then the first real dictation would land empty with no obvious reason why.

If you ever saw HyperVoice land on a full-page "Loading speech engine…" spinner that wouldn't go away, that's fixed. The fix is a small one-liner but it removes the case entirely so you should never see a stuck loader again on launch.

A second pass on the new Parakeet speech engine added in v0.5.134.

- DirectML GPU acceleration — Parakeet can now run on any DirectX 12-capable GPU (NVIDIA, AMD, Intel). Off by default because on the int8 quantized model, modern CPUs are often just as fast or faster for dictation-length audio. Toggle it under Settings → AI → GPU Acceleration when Parakeet is the active engine.

- Whisper and Parakeet now have independent GPU preferences. If Whisper's Vulkan crashed once on your machine and auto-disabled itself, that no longer cascades into disabling Parakeet too. Each engine remembers its own setting.

- Parakeet is now labelled "Experimental" in the engine dropdown, the model card, and the confirm dialog. Accuracy and speed may still differ from Whisper — switching back and forth is free once both models are downloaded.

- Status bar model popup combines both engines into one list now: a WHISPER section listing your downloaded Whisper variants and a PARAKEET section listing downloaded Parakeet variants. Click any row to switch engine + variant in one move.

- First dictation after launch is fast for Parakeet users now — the engine preloads in the background at app launch (previously only on engine-selector clicks), so you don't pay a ~1.5 second cold-start when you press your hotkey for the first time.

- Fixed: the status bar model name (e.g. "Tiny English" or "Parakeet TDT 0.6B v3") and the active checkmark in the popup are now in sync with each other and with Settings → AI. Previously two of those three surfaces could drift apart when you switched engines from one place vs another.

- The By Model breakdown on your Stats page now shows the average time each speech model takes on your machine alongside the existing transcription count and word total. Each row now reads N transcriptions · M words · avg X.XXs (or XXms when sub-second).

- Useful for deciding whether Parakeet or one of the Whisper variants is the better daily driver for your hardware — the numbers are computed from your own dictation history, not a synthetic benchmark.

- Outlier guard: stuck-hotkey records and recordings longer than 60 seconds are excluded from the average so a single edge case can't skew the model's number.

- You can now choose between Whisper and Parakeet for speech recognition. Go to Settings → AI → Speech recognition engine to switch. Whisper stays the default for every existing install; nothing changes unless you opt in.

- Parakeet TDT 0.6B v3 is NVIDIA's CPU-optimised speech model — well suited for laptops without a discrete GPU, machines where Whisper's GPU path has been flaky, or anyone who prefers a smaller resident memory footprint at idle.

- Multilingual auto-detection across 25 European languages (English, Spanish, French, German, Italian, Portuguese, Dutch, Polish, Czech, Romanian, Greek, Swedish, Danish, Hungarian, Finnish, Bulgarian, Croatian, Slovak, Slovenian, Lithuanian, Latvian, Estonian, Maltese, Ukrainian, Russian). For languages outside that set (Japanese, Mandarin, Arabic, Hindi, Korean, etc.) Whisper is still the right choice.

- Picking Parakeet kicks off a one-time ~670 MB download with a confirm dialog so you opt in deliberately. After that it runs fully offline, exactly like Whisper. Both engines stay on-device — neither sends your speech to the cloud.

- Switching back and forth is free once the model is downloaded — no second confirmation, no re-download. You can keep Whisper as your daily driver and try Parakeet on a tough audio clip with two clicks.

- AI post-processing (HyperVoice Cloud / OpenAI / Anthropic) works the same regardless of which engine you pick — it operates on the transcript text after speech recognition finishes.

- First-visit banner now asks for your choice on analytics. Click Accept all to keep things as they were (we use Google Analytics and Microsoft Clarity to understand what's working on the site); click Reject analytics and neither tool gets a cross-page identifier from you — GA falls back to anonymous aggregate counters and Clarity doesn't load at all.

- Necessary cookies (your sign-in session, your saved theme) are unaffected either way.

- A Cookie preferences link in the footer lets you change your mind at any time.

- Full details in section 2.5 and the dated update log at the bottom of the Privacy Policy.

- If you sign in on a non-Windows device, you're now automatically on the waitlist for that platform instead of having to click a "Notify me" button. The previous opt-in design was missing nearly everyone — most non-Windows visitors saw the "Windows-only" banner and bounced before noticing the button. Now the banner just confirms enrolment and gives you a one-click "Remove me from the list" if you don't want it.

- Removal sticks per-browser — if you opt out, we won't re-enrol you on subsequent visits.

- One email per platform launch, no marketing pings. The waitlist exists to email you the moment a build for your platform is ready — that's it.

- Skipping the onboarding mic test now asks you to confirm. Previously it was a small button next to the big Next — easy to click without reading. Now you get a quick "are you sure?" with a default "Test it now" because skipping is the most common reason new users hit silent first dictations and give up.

- You'll now see a banner when HyperVoice records but no audio comes through. Before this, a recording that captured zero samples (mic permission not granted, hardware mute toggle on, wrong default device) ended with a flash toast that vanished in 5 seconds. Now it's a sticky banner explaining the three common causes in order — Windows microphone permission, hardware mute switch, wrong default mic — with a one-click link to Sound Settings. The banner clears itself the moment your next real dictation lands.

- Settings tabs reorganised so things you fiddle with together live in the same place. New layout:

- General: theme, launch at startup, streamer mode (owner-only).

- Dictation: the test-it-here textbox, microphone, hotkey, recording mode (toggle vs push-to-talk), output mode (auto-paste vs clipboard).

- AI: speech models, GPU acceleration, post-processing provider.

- Account and About unchanged.

- Old "Audio" tab is gone — its mic picker moved to Dictation and the GPU toggle moved to AI (where it actually belongs).

- Same number of tabs, just grouped so you don't bounce across three tabs to fix one thing.

- Fix: sometimes the wrong thing pasted (your previous clipboard contents instead of your dictation). HyperVoice puts your transcription on the clipboard, sends Ctrl+V, waits briefly, and restores whatever you'd had on the clipboard before. The wait was too short for some slower target apps (Win11 Notepad, Windows Terminal / Claude Code, Slack, VS Code, Discord) — the app was still mid-read when we restored, so it ended up pasting the restored-previous-contents instead of your dictation. Bumped the wait from 80ms to 250ms (still well below "feels laggy") and the problem goes away.

- The floating pill now shows errors in place when the main HyperVoice window is minimised. Previously, things like "the target app is running as administrator, paste blocked" only appeared inside the main window's notification banner — which you couldn't see if you were dictating into a different app. Now the pill morphs into a small amber alert with a short label (e.g. "Admin app — Ctrl+V to paste") right where your eyes already are. Click the pill to open HyperVoice for the full explanation, or the × to dismiss. Starting a fresh dictation also clears it automatically.

- HyperVoice now tells you when it can't paste because the target app is running as administrator. Windows blocks any non-admin app from sending keystrokes into an admin-elevated window — that's a security feature, not something we can override. Before this release the dictation would silently land nowhere (your text was on the clipboard but the Ctrl+V never reached the target). Now you'll see a banner: "Can't paste into an app running as administrator — your text is on the clipboard, press Ctrl+V to paste it manually. To fix it for next time, restart the target app without 'Run as administrator'." Most common case: command prompts opened with Run as administrator.

- Fix: pastes silently disappearing into Claude Code, Windows Terminal, Win11 Notepad, and similar modern apps. v0.5.125 traded a settle delay away to fix a different issue, which accidentally broke paste reliability in apps that use Windows' newer text framework (Cascadia / WinAppSDK). The delay is back — it's only ~15ms, you won't notice it — and your dictations now land in those apps as expected.

- Free-tier post-processing tidied up. If your trial or paid plan expires and you'd previously picked a cloud post-processing mode (Clean Up, Professional Email, etc.), HyperVoice now resets your processing mode and provider back to None automatically and lets you know with a toast. Before this, the pipeline was already skipping the cloud call when your plan no longer covered it, but Settings → Processing still looked like your old mode was active — confusing if you noticed the mismatch. If you later upgrade again, just re-pick the mode you want.

- Fix: trial-expired accounts no longer get stuck showing "Trial: -55 days remaining" in the title bar. A small group of accounts whose 7-day trial ended a while ago weren't being transitioned to the free tier — the desktop kept reading them as on-trial and counting backwards from the expiry date. The server now flips the status the moment the trial date passes, and the desktop title bar reflects that correctly.

- Fix: dictations sometimes pasted a literal "v" instead of your text. After v0.5.124, a small timing race could let your hotkey release event slip in between our injected Ctrl-press and V-press, leaving the target app to see V on its own — and a stray "v" would land instead of your sentence. Rewrote the paste keystroke to send Ctrl+V as one atomic burst (the OS guarantees nothing else gets in the way), and tested it through ~40 back-to-back pastes with no recurrence.

- Clearer warning when your hotkey contains Alt. The wording in onboarding and Settings → Hotkey was pretty technical. Now it just says: "In Windows apps like Notepad, Word, and Excel, pressing Alt opens the menu bar. If your hotkey contains Alt, pasted text can vanish into a menu instead of landing in your document. Pick a combo without Alt to avoid this."

- Fix: dictating into Notepad sometimes pasted nothing. HyperVoice was transcribing and putting the text on the clipboard correctly, but Ctrl+V was being sent a fraction of a second before Notepad's text control was ready, so the keystroke vanished. A tiny extra delay (15ms — invisible to you) gives the target app time to be ready, and Notepad now reliably accepts the paste. The same fix helps Word, Excel, and other Win11 apps that use the modern text control.

- Warning if your hotkey contains Alt. In Settings → General → Hotkey (and during onboarding), if your chosen combo includes the Alt key, you'll now see an amber heads-up: Alt-based hotkeys conflict with the menu access keys in Notepad / Word / Excel / Outlook and pasted text may be silently dropped into those apps. If you hit this, change to a combo without Alt — Ctrl+Shift+Space is the default and works everywhere.

- The aurora animation now also plays inside the app. When you press the hotkey while HyperVoice is the active window, the recording bar that drops down from the top of the window used to show a small row of bars — it now shows the same flowing aurora you already see on the floating pill and in the microphone test during onboarding. One visual language across every place HyperVoice listens to you.

- The floating pill has a fresh look. While you're dictating, flowing bands of brand colour now ripple through the pill in real time — louder voice = brighter aurora, quiet = ghostly. The previous 16-bar waveform was technically responsive but read as a near-static row of orange dots to most people; the new visualisation gives a much clearer "yes, we hear you" signal at a glance. Same hotkey, same workflow — just a more alive look.

- The pill is a touch larger so the timer on the right is no longer clipped against the rounded edge.

- Right-click the pill → Quit HyperVoice. Until now you had to find the system tray icon to fully exit the app. Now Quit lives in the same menu as Record / Settings / Show HyperVoice.

- The microphone test in onboarding shows the same aurora, scaled up. The first time you press the hotkey for real, the visual is one you've already seen.

- New onboarding step asks you to try the hotkey in a real app before finishing setup. After the "How it works" step you'll see a "Try it for real" screen: open Notepad (or your browser, or an email draft), press your hotkey, and watch your words appear there. Then come back and click "I tried it." We added this because we noticed a lot of people were finishing onboarding without ever actually using the hotkey — and then never used the app again because the muscle memory hadn't formed. One real-world dictation while you're still in setup mode makes a huge difference.

- Persistent "ready to dictate" hint at the top of the main window if you've installed HyperVoice but haven't dictated yet. After 24 hours of inactivity it shifts to a "Stuck? Tell us what's blocking you" message with a one-click link to the Feedback page. Vanishes the moment your first dictation lands.

- New "Try dictation here" box on the Settings page. A focused text area at the top of Settings — click into it, press your hotkey, say something, and the transcript lands right there. No need to leave HyperVoice to test that it's working.

- Diagnostic banners now stay on screen until you dismiss them. v0.5.119 introduced the persistent banner format, but a separate auto-hide on the recording overlay was clearing the banner along with the overlay state about 1.5 seconds after a dictation finished. So a "We barely picked anything up" message you were meant to read for 10–30 seconds was visibly flashing on screen for ~1–2s and disappearing. Now the banner stays until you either click the X, or your next successful dictation auto-clears it.

- Error toasts now show for the full 5 seconds. Same root cause: the recording overlay's auto-hide was wiping error messages too at the 3-second mark, even though the toast is meant to last 5 seconds. Affected messages like "AI processing failed: rate limit exceeded, retry in 60s" — useful info you'd miss by the time you read it. Now the toast's own timer drives its lifetime.

- Diagnostic messages now stay on screen until you dismiss them. v0.5.118 used the same 5-second toast as transient errors — too short to read a multi-sentence explanation and decide what to do. Now diagnostics live in a persistent banner at the top of the HyperVoice window, sitting alongside the other notification banners. It auto-clears the moment your next dictation succeeds, or you can dismiss it with the X.

- Each banner carries an "Open Sound Settings" button and a "Get help" menu. Get help opens Send feedback (the in-app feedback page) or Join us on Discord (the community where you can ask in real time).

- Three new diagnoses surface here:

- No microphone detected — fires the moment HyperVoice tries to record but Windows reports zero input devices. Previously this was a silent failure (the recording just didn't happen). Now you see the device name HyperVoice tried, plus the same Sound Settings shortcut.

- Audio capture didn't keep up — fires when the wallclock duration of your hotkey press is much longer than the audio that was actually captured (e.g. you held the hotkey for 11 seconds but only 0.9 seconds of audio came back). Almost always means another app — Teams, Zoom, Discord, OBS, a meeting client — grabbed the microphone mid-recording. Banner tells you to close those apps and retry.

- Onboarding mic-test step now blocks if you have zero microphones. If Windows reports no input devices when you reach the mic-test step, the test/skip buttons are replaced with a clear "Connect a microphone, then Refresh" panel — same Open Sound Settings shortcut, plus a Refresh button so you can plug in a mic without restarting onboarding. You can still continue without one, but it's no longer a silent breeze-through.

- You now get a pinpointed message when a dictation comes back empty. Previously, if you held the hotkey, said something, but nothing was pasted, the app went silent — no error, no hint, just an empty result that left you wondering whether to retry or give up. Now we run the same diagnosis the v0.5.117 mic test runs, but on real dictations:

- Captured audio but it was barely any sound → "Captured Xs of audio but barely any sound — your mic might be muted at the hardware switch, or the wrong device is selected. Switch microphones in Settings → Audio."

- Captured audio at normal volume but no recognisable words → "Captured Xs of audio but couldn't make out any words — try speaking a bit louder and more clearly, or switch to a different mic in Settings → Audio."

- Why this matters: users who skipped the onboarding mic test (about half of them, per the deliberate "Skip test" flow added in v0.5.116) had no diagnostic surface at all when their first real attempt came up empty. The classifier was already in the code — it just only fired on the test step. Now it fires wherever it's needed.

- If you sign in on macOS, iOS, Android, or Linux, the dashboard now offers a one-click "Notify me when {your platform} ships" button instead of asking you to compose a support email. Re-visit the dashboard later and it'll show "You're on the {platform} waitlist" so you don't double-enrol or wonder if it took. Email is the one already on file (no re-typing), and we'll send a single message when there's a build for you to install — no marketing pings in between.

- **The mic test now tells you why it didn't work, not just that it didn't work.** Previously when the test came back empty, you got a generic "we didn't pick anything up — try again" message that left you guessing what to do. Now we look at what we actually captured and show one of three targeted hints:

- No audio at all → "Your mic might be muted, or another app could be using it (Teams, Zoom, Discord, OBS). Try a different microphone above…"

- Audio but very quiet → "We captured a few seconds of audio, but barely any sound. Likely the mute button on your headset is on, or a different mic is selected as default…"

- Audio but no recognisable words → "The volume looked fine, but no recognisable speech came through. Try speaking a bit louder and more clearly…"

- Sound settings open in one click. When the test comes back empty, a new "Open Sound settings" button takes you straight to Windows' sound input settings — no more digging through Settings menus to figure out which mic is set as default.

- The microphone picker briefly highlights itself with a subtle blue pulse if the test came back empty — a visual nudge that swapping mics is the most useful next step.

- Onboarding mic test now requires a deliberate "Skip test" if you don't run it. Previously the same shiny brand-coloured "Next" button sat there at every step, so it was easy to breeze past the mic test without ever pressing record — and then discover the mic wasn't working three days later when your first dictation came up empty. Now the Next button on the mic-test step changes to a quieter outlined "Skip test" button until you've actually completed a test (success or failure). You can still skip — but you have to mean it. Helps catch wrong-default-mic problems during setup, where they're cheapest to fix.

- You can now find larger speech models more easily. The status-bar model dropdown's "More models..." link is now "Download a more accurate model" with a one-line note underneath: Larger models are more accurate but slower. Same place — bottom of the dropdown when you click the model name in the status bar.

- Fixed: that link used to take you to Settings → General. It now correctly opens Settings → Processing, where the speech-model picker actually lives.

- Onboarding now ends with a hint about model upgrades. The Summary step has a small line at the bottom: "Want better accuracy later? Settings → Processing has larger speech models you can download anytime — they're more accurate but slower." So a new user who finds the default Base model's accuracy underwhelming knows there's a path to better without having to ask.

- The dashboard tab bar now has a Download button anchored to the right. Visible regardless of which tab you're on, so when you want to install HyperVoice on a second machine you can grab the installer in one click instead of having to know to dig into the Devices tab. Hover the button for the "Install on another device" tooltip.

- Why: previously, the big animated "Download for Windows" button on the Get Started tab disappeared as soon as you'd downloaded once — sensible for first-time users (don't keep nagging) but it left returning users with no obvious download path. The only existing fallback was a small inline link buried in the Devices tab. Now the option is always one click away.

- Fixed: home screen looked like it was recording right after onboarding. When you finished the onboarding microphone test and the app moved to the main view, the recording overlay would paint over the home screen as if you were currently dictating — even though you weren't, and even though the floating pill was correctly back to idle. Cause: the test recording entered the "recording" state but never had a matching "I'm done" event the main window could hear, so the global status was stuck. Now resets cleanly the moment the test transcription completes.

- Microphone test step has a live waveform now. While you're holding the recording during the onboarding mic test, a 16-bar audio-level visualizer pulses with your voice — the same one in the floating pill. Confirms the mic is hot before you commit to reading the script (and stops the "is it actually working?" doubt that came from a static red button).

- Pick your microphone right inside the test step. A small "Microphone: ▼" dropdown sits between the test sentence and the start button. Most "we didn't pick anything up" failures are wrong-default-mic problems (laptop internal vs USB headset vs conference dongle), and now you can switch without leaving onboarding to dig through Settings. Defaults to your system default mic, lists all detected input devices.

- Fixed: pill stuck on "Transcribing..." after the test. When you finished the mic test, the floating pill window stayed in the "Transcribing..." state forever instead of returning to idle — the test transcription was missing the "I'm done" signal that the pill needed. Now resets cleanly.

- Onboarding got a much friendlier first run. The old "pick a speech model" step was a question almost no one knew how to answer ("Tiny? Base? What's a Whisper Medium?"). It's gone. We now ask one question — what language do you speak — pick the right speech model for you, download it with a clean progress animation, and immediately put you through a microphone test so you know everything works before the rest of setup.

- The microphone test gives you a sentence to read in your own language, then transcribes it on the spot and shows what we heard. If we got something, you see "Looks great — your microphone is working." If we didn't pick anything up, you get a clear "We didn't pick anything up — try again." Press the big start/stop button (no hotkey needed yet), short and simple. Localized to the 10 most common HyperVoice languages.

- Multilingual support is now a first-class onboarding choice. Spanish, French, German, Portuguese, Italian, Dutch, Japanese, Mandarin, Hindi — pick one and the right model is downloaded automatically. There's also an "Other (multilingual)" option for any other language. Nothing about the rest of the app is translated yet (still English UI), but transcription now actually uses the language you chose instead of always assuming English under the hood.

- Onboarding "Next" no longer throws away an unsaved hotkey. If you click Change on the hotkey step, type a new combination, and then click Next instead of Save, the new combo is now saved automatically (rather than silently discarded as before). Captured + valid → saved + advanced. Captured + conflicting with another app → stays on the step with the inline error so you can pick another. Clicked Change but never typed anything → the "change" attempt is dismissed and you advance with whatever was already there.

- "Sign In with HyperVoice" is now "Sign in to HyperVoice" — small grammar fix on the account-link screen + matching dashboard instructions. The previous wording followed the "Sign in with Google" convention used when X is an external auth provider, which made the desktop button read like HyperVoice was logging you into something else.

- Friendly recovery flow when your keyboard shortcut is taken. If another app on your computer is already using Ctrl+Shift+Space (the default), HyperVoice now opens a focused "Pick a keyboard shortcut" sheet instead of leaving you with a tiny error toast and a non-working app. The sheet shows the failed combo as physical-looking key chips, lets you press a new combination and see live whether it's available (green tick) or also in use (red), and offers three one-click suggestions of combos that are usually free (Ctrl+Alt+Space, Ctrl+Shift+H, Alt+Shift+V). Press Enter to save, Esc to skip.

- Skip-for-now path stays beautiful too. If you'd rather decide later, a slim "No keyboard shortcut active" banner stays at the top of the app with a one-click "Choose shortcut" button — so you can come back to it whenever you want without having to dig through Settings.

- The recovery sheet now also appears during onboarding. If your default hotkey is taken on a fresh install, you'll be prompted to pick a working one as part of the first-run flow rather than after — no more "I picked Continue and now there's a surprise modal" moment.

- New "Avg time per dictation" chart on your Statistics page. Shows, day by day for the last 30 days, how long each of your dictations took to process end-to-end (Whisper transcription plus any post-processing), in milliseconds. Lower is faster.

- Toggle the split between All, GPU vs CPU, and Whisper model to see whether your speed is being driven by your hardware, the model size you've picked, or just normal day-to-day variance. Useful before deciding "is it worth downloading the larger model?" or "is the GPU path actually helping me?"

- Section 2.4 of the Privacy Policy has been rewritten to spell out exactly what we capture. Where it previously said something vague like "we may collect anonymized aggregated usage data," it now lists the actual categories — machine metadata (CPU, RAM, mics, OS), reliability events (crashes, hotkey/audio failures, pipeline errors), feature usage (provider/mode + a couple of UI events), website errors, and a single platform bucket inferred from your browser.

- A new "What we never collect" paragraph spells out the inverse: the content of what you dictate or transcribe, any character/word count or hash derived from that text, form inputs, cookies, localStorage, keystrokes outside the recording hotkey, files on your machine, browsing data.

- Nothing about what we actually capture has changed — every concrete capture was already documented entry-by-entry in the Data & Privacy Update Log at the bottom of the policy. This rewrite just brings the prose at the top of the page in line with the log, so the policy itself reads accurately without you needing to scroll to section 10.

- Your Statistics cards now show trends at a glance. The Transcriptions, Total Words, and Speaking Time cards have a small 14-day sparkline along the bottom, so you can see how you've been trending without scrolling down to the full chart. An upward curve usually means a good week.

- Avg Speed now shows how this week compares to last. Because average speaking speed is a ratio, a tiny sparkline would swing wildly on a quiet day. Instead the card shows something like ↑ 12% vs previous 7d — green arrow up if you're getting faster, red arrow down if slower, "flat" if you're within a couple of percent. Hover for the exact numbers. It stays hidden if either week didn't have enough dictation for a fair comparison.

- Account linking no longer asks for a 6-character setup code. Click "Sign In with HyperVoice", sign in in your browser, and the app links automatically — that's the whole flow. You'll no longer see a code-entry box in the app, and the dashboard no longer has a "Browser flow didn't work? Use a setup code instead" panel to hunt for.

- Why the change: the code was a fallback for the rare case where your browser couldn't hand the sign-in back to the desktop app, but its presence in the UI made the main sign-in look less trustworthy than it is. The app now does that fallback work invisibly — if the direct hand-off doesn't complete within a couple of seconds, the app finishes the link itself by checking with our server in the background. You never see anything fail, and you never type a code.

- If the browser doesn't open at all, the "copy this URL into any browser" fallback still works — paste it anywhere and signing in on that browser will still link the app automatically.

- Auto-update no longer interrupts a model download. If HyperVoice was downloading a speech model in the background and a new version finished installing at the same time, the app used to immediately restart to apply the update — killing your download halfway and wasting the bandwidth. Now the update waits politely for the model download to finish before restarting. You'll see a small "Update ready — restarting after {model} finishes downloading…" banner so you know what's going on.

- Better diagnostics for setup problems. When you launch HyperVoice we now check whether your machine actually has what's needed (CPU, RAM, microphone) and report the result back so we can spot when someone's hardware can't run the app well. We also detect when key things go wrong silently — like your hotkey already being claimed by another app, or there being no microphone available — instead of just leaving you wondering why nothing's happening.

- This is purely about us being able to diagnose and fix issues faster. See /privacy → Data & Privacy Update Log for exactly what's captured (machine metadata only — never what you say or type).

- Every change that affects what we store or transmit is now listed in one place. New "Data & Privacy Update Log" section on /privacy, reverse-chronological, with "What's captured" and "What's NOT captured" spelled out for every entry. A prominent "What's changed recently" chip near the top of the policy page takes you straight there.

- Going forward, we stop adding "Privacy note" boilerplate to every changelog entry — dilution makes those notes easier to skim past. The dedicated log is the authoritative, readable-in-one-sitting source. Purely cosmetic / UI changes are kept out on purpose — only data-handling changes go there.

- Per-dictation WPM on the History tab. Every row now shows how fast you spoke for that specific dictation — green + bold when you beat your own 7-day average, neutral otherwise. Hover for the exact number and what you're being compared against. Combined with yesterday's voiced-time fix, these numbers are accurate even on long-hold dictations where you paused partway.

- Avg Speed (WPM) is now accurate again. Previously, if you accidentally held the hotkey long after you stopped speaking, those seconds of silence got counted as speaking time and dragged your average speed down. We now use Whisper's own segment timestamps to measure voiced time only — so a 4-second hotkey press with 2.5s of speech counts as 2.5s, not 4s. Plus we filter out obvious stuck-hotkey outliers (over 15 seconds of audio with fewer than 30 WPM worth of words — almost always accidental). Your existing history is included in the filter immediately; new dictations from this version onward also benefit from the more accurate voiced-time measurement.

- Sign-in now recovers when your browser doesn't open. If you clicked "Sign In with HyperVoice" on the account link screen and Windows silently cancelled opening your browser (a UAC prompt you dismissed, no default browser set, or antivirus blocking), the app used to show a cryptic error and leave you stuck. Now it shows the exact sign-in URL with a Copy button — paste it into any browser manually and the flow completes the same way.

- Non-Windows visitors get a heads-up on the Dashboard. If you signed up from a Mac, iPhone, Android, or Linux machine, the dashboard now tells you clearly that HyperVoice is Windows-only right now (macOS on the roadmap) with a waitlist email link — instead of silently letting you download an installer that won't run on your OS.

- Better diagnostics when something fails. If your dictation hits a processing error, we now capture the upstream request ID from the provider (OpenAI / Anthropic / our Cloud) alongside the error message. Makes it much faster for us to trace a specific failure end-to-end and — if needed — raise a support ticket with the provider. Zero user-visible change.

- HyperVoice Cloud post-processing now works if you signed in via the new account flow. If you were getting "Cloud error: Not authenticated" when using Clean Up, Summarize, Professional Email, etc., this release fixes it. The cause was a gap in the auth wiring for users who signed in via their HyperVoice account (rather than pasting a license key) — the Cloud call wasn't sending your session. Fixed on both the desktop and the server. Your dictation falls back to raw text on older versions, so nothing was lost — but you'll now get the actual processed output.

- More complete error reporting for us. Extending yesterday's change so all pipeline failures — Whisper errors, audio issues, paste problems, anything that previously just aborted silently — now get reported back to us. Nothing changes for you; this is about making sure if something breaks for a user, we can actually see it and fix it.

- "What's new" signals added to the website. A small pill on the landing page shows when HyperVoice was last updated. Logged in? Your dashboard now has a small "What's new" link above the tabs with a pulsing dot when there's a release you haven't seen yet — mirrors the bell panel in the desktop app.

- Post-processing failures are now reported back to us. Previously if a processing mode errored (e.g. a BYOK API key was missing, a provider was unreachable, or the LLM returned an error), the app silently fell back to the raw transcription and swallowed the error — making it impossible for us to spot what went wrong for users. Now processing failures are logged centrally so we can see them, debug them, and fix them without waiting for users to report. Your dictation still falls back to raw text, so nothing you do changes — the change is purely for visibility on our end.

- Settings → About → Updates now has a single always-visible "Check for Updates" button. Previously the button disappeared once you'd checked and was replaced by a smaller "Check again" link in the up-to-date state. Now the button stays put on the left in every state, and the status — "You're on the latest version," "Checking…," "v0.5.94 is available," an error, etc. — sits to its right. Re-checking is always a single click from where your eye last saw the button.

- Dictionary replacements no longer leave a trailing period on slash-command-style rules. If you've set up a replacement like clear command → /clear and you speak the phrase on its own, Whisper adds a period to the end of the transcript ("Clear command.") before the replacement runs. The old behavior left you with /clear. in the paste. Now, when your entire spoken phrase is the single replacement trigger, HyperVoice returns just the replacement value — no stray punctuation. Multi-sentence dictations ("Clear command. Then save.") still keep their punctuation unchanged.

- Discord card above the tabs. Whichever dashboard tab you're on (Get Started, Subscription, Devices, Feedback, Account), a "Need help or have a question?" Discord card sits in the persistent area at the top. Sits directly below the download banner when that's showing, so first-time users hitting install friction see the community support path immediately.

- Fixed a stale "You haven't downloaded HyperVoice yet" banner. It used to linger for users who had downloaded and were actively dictating, if they'd cleared cookies, signed in fresh on a new device, or installed via a different browser. Now the banner respects the server-side record of your install + usage, so if the server knows you're installed the banner hides. Also writes the signal back to your local browser so the banner doesn't flash on your next visit either.

- "You're all set up" card on Get Started is now dismissible. Once you've seen the green confirmation, click the × in the top-right to hide it across sessions. Your stats below stay visible as normal. If you ever want it back, clear the hv_all_set_dismissed localStorage key.

- The Get Started tab still has its own bigger Discord card inside the checklist, for prominence at the moment of setup.

- There's a Discord now. A place to ask questions, report bugs, and share what you're building with voice input. Usually faster than email — the team and other users are there live.

- Top nav: "Community" link in the main menu (next to Features, Pricing, Blog) and a Discord icon chip next to the theme toggle, on every page.

- On the web: every footer has a Discord link. The dashboard's "Get Started" checklist has a prominent "Stuck? Ask in Discord" card that points at #install-help for install-related questions.

- In the desktop app: Settings → About has a "Join the Discord" card. The Feedback tab suggests Discord for anything that benefits from a back-and-forth. "What's New" ends with a "Discuss this release" link.

- Invite (unlimited, never expires): discord.gg/EM5sFnCUCX.

- New "What's New" in the sidebar. A bell icon between Feedback and Settings opens this panel. You'll see a small dot on it whenever there's something new; clicking the panel clears it.

- Remotely unlinking a device now works properly. If you unlink this machine from your web dashboard, HyperVoice here drops back to the Account Link screen on its next check (up to an hour, or immediately on restart). Before this you could end up in a broken in-between state.

- Expired or renewed subscriptions now sync automatically. The app re-checks your plan from the server every hour, so you don't need to restart after a renewal, upgrade, or cancellation.

- Clearer banner when a Pro subscription lapses. If your Pro plan ends and you drop to the free tier, you'll see a distinct amber banner with a "Reactivate" link — separate from the regular free-tier banner.

- Cancelled or failed-payment subscriptions keep Pro access through the period you paid for. You won't get cut off mid-month for a failed renewal while the payment is being retried.

- Fixed a flickering/reload loop on the Subscription page for users on the newer account-linked flow.

- Subscription page shows your plan details correctly for account-linked users — plan, status, renewal date, all populated from the server, with a Refresh button.

- You no longer deal with license keys. Your plan syncs automatically from your HyperVoice web account — sign in once and everything just works. Existing users with a license key keep working until they link their account; after linking, the old key is retired.

- Offline grace period — if you're away from the internet, HyperVoice keeps working for 14 days on whatever your last known plan was, rather than locking you out.

- Safer settings saves. If HyperVoice or your machine crashed while saving settings, the file could sometimes become corrupted and prevent startup. That can't happen any more.

- More reliable stop → start in quick succession. Pressing your hotkey rapidly no longer risks the old recording bleeding into the new one.

- Pasting to the right app, every time. If the window you originally recorded into has been closed and its window handle recycled to another app, HyperVoice now detects that and skips the paste rather than pasting into the wrong place.

- Faster paste on snappy systems, more reliable on busy ones. Replaced a fixed wait with an active poll so paste happens as soon as the target window is ready.

- GPU stays enabled after harmless errors. Previously a single failed model load could disable GPU permanently across restarts. Only an actual crash in the GPU code now disables it (the original intent).

- Account Link screen recovery paths — 3-minute timeout, "still waiting?" helper after 15 seconds, one-click retry / manual-code / open-dashboard buttons when something goes wrong. Friendlier messages for expired or wrong setup codes.

- Onboarding tells you what the app does and how long setup takes before asking you to do anything. Model downloads now show bytes-so-far and a rolling ETA.

- Faster provider key lookups — checks an in-memory cache before hitting the OS keyring. You'll notice this whenever HyperVoice starts a dictation.

- Your stats, now with charts. Get Started's "You're all set up" screen now shows total dictations, words transcribed, estimated time saved, daily-activity chart, day-of-week breakdown, current streak, and average latency.

- New Devices tab. See every linked device (name, last-active, app version, first-linked date) and unlink any of them remotely. Device limit is shown up-front.

- Better Subscription tab. Renamed from "Plan", handles every subscription state — active / trial / cancelled-with-access / past-due / lapsed — with clear next-step CTAs.

- Get Started checklist. 4-step live progress bar (create account → download → link → dictate) that updates as you go, with a celebratory "You're all set up" screen when done.

- Display name pulled from Google sign-in. Previously every Google signup showed blank name in the dashboard; now your name is picked up automatically.

- Timestamps show your local time. "Last active 2h ago" is now accurate regardless of your timezone.

- New walkthrough at hypervoice.app/install-help explaining Windows' "protected your PC" warning with step-by-step screenshots and a plain-English "why this happens" explainer.

- Linked from the dashboard at the moment you click Download, so the help is there if you need it.

- HyperVoice is now free — 500 words/day, forever, no credit card. Pricing, homepage, and FAQ reflect it.

- Clearer Pro vs Lifetime framing. Pricing section explains the upgrade trigger ("when 500 words/day isn't enough, or when raw dictation needs polish before it ships") rather than just listing features.

- Polished copy throughout — CTAs say what they actually do (no more "Sign Up to Download" chained verbs), warnings on the homepage explain themselves, pricing subtitles reflect real benefits.

- Hero loads faster. The big headline on the homepage now paints instantly instead of fading in after 700ms.

- Respects "reduce motion" system setting. If you've turned off animations at the OS level, HyperVoice's website honours it everywhere.

- Friendly auth errors. Sign-in / sign-up errors now read like English ("That email address doesn't look right. Check for typos.") instead of Firebase jargon.

- Forgot password link on the sign-in form with a one-click password-reset email.

- Password strength meter on sign-up so you can see if your password is weak before submitting.

## Older releases

This public changelog started with v0.5.87. For questions about earlier builds, email support@hypervoice.app or ask in #install-help on Discord.