VoiceWave Local Core: baixar versions anteriors

The privacy check now reaches people who start VoiceWave with Windows, and
nothing is lost while you decide.
Changed
- Your reports wait while you decide. If you shared the older usage
count before 0.7.7 and haven't answered "A quick privacy check" yet,
VoiceWave now keeps the anonymous reports on your PC instead of
skipping them. Nothing is sent until you answer. Keep sharing sends
them; Turn off deletes them and sends nothing.
- The privacy check finds you. If VoiceWave starts hidden with
Windows, the check could wait unseen. Now, after your first dictation
that works in each launch, the pill asks once, and Open shows it.
- Saying yes is counted. Choosing to share, at the end of setup or on
the privacy check, is now one anonymous report. Choosing not to share
sends nothing.

Better problem reports, so we can find and fix what goes wrong for you, and
a support file you can attach to an issue.
Added
- Something wrong? in Help → Report a problem. Pick what happened, such
as wrong words, too slow, or the text didn't appear. If you share
anonymous reports, we get the reason you picked and nothing else. Then
tell us more on GitHub, with the reason already filled in, or save a
support file to attach.
- Save support file. Settings → Diagnostics → Save support file writes
a zip to your Downloads folder with your diagnostics report, the newest
logs, and how the last run ended if VoiceWave closed unexpectedly. Your
Windows user folder is hidden in it, and it is never uploaded. Show in
folder finds it in Explorer.
- A note after VoiceWave closes unexpectedly. If the last run crashed
or hit an internal error, Home says so and offers Save support file and
Open logs folder, until you dismiss it.
- Where did you hear about VoiceWave? One optional question at the end
of setup on a new install, asked once and only while anonymous reports
are on.
Changed
- Anonymous usage and problem reports move to PostHog. They go to
PostHog's EU Cloud, in a project set to discard IP addresses, and now say
how VoiceWave works for you: one summary for each day you dictate (how
many dictations worked or failed, how fast text appeared, and the kind of
app it went into, never its name), how the last run ended, model
downloads, failed updates and crashes. The full list is under What's sent
in the app and in the
privacy notice.
Never audio, transcripts, app names, file paths or your IP address.
- Asked once, one switch. If you shared the older usage count, a quick
privacy check asks once whether to keep sharing, and nothing new is
recorded until you answer. If you had it off, you aren't asked and
nothing is sent. One switch in Settings → Diagnostics turns everything
off and deletes anything waiting to be sent.
- Reports wait for a connection. Reports made while you're offline are
kept on your PC for up to 14 days and sent later with the time they
happened, instead of being lost.
- TelemetryDeck for one more release. 0.7.7 still sends the older
anonymous usage count through TelemetryDeck, under the same switch. That
stops in the next release.
Fixed
- The close button works after reopening VoiceWave. Since 0.7.5, if
you opened VoiceWave from the Start menu or a shortcut while it was
already running, the window's close button stopped sending it to the
tray. It hides to the tray every time again.
- A failed update says so plainly. Instead of a raw error, the update
prompt and Settings → Updates now say the update didn't install, that
VoiceWave will try again later, and that you can download the latest
version from voicewave.site.

Speech model downloads now survive a bad connection, and say what went wrong
when they can't.
Fixed
- Speech model downloads survive a dropped connection. If the network
stalls or drops (for example when the PC sleeps), the download waits a
moment and picks up from where it stopped, for up to about two minutes
without progress, which covers a laptop waking from sleep. It also works
with antivirus or company networks that check secure connections with
their own certificate.
- A failed download says why. Instead of a generic error, you see what
went wrong (no connection, blocked secure connection, disk full, file held
by another program) and what to do next.

Parakeet becomes the default speech model, with your dictionary words built in. Also dictation into terminals, "press enter" to send, Try again after a failed transcription, and an easy way to report problems.
New
- Parakeet, the new default speech model. NVIDIA's Parakeet TDT 0.6B v3 is about 3 to 4 times faster than Large Turbo on a graphics card, faster than Small on a processor, and returns nothing instead of made-up words when you say nothing. It understands English and 24 European languages. Existing users see a one-time card on Home to switch; your current model keeps working until Parakeet is downloaded, and Small and Large Turbo stay in Models. Parakeet is by NVIDIA, licensed under CC BY 4.0.
- Your dictionary words steer Parakeet, so names like "Claude" come out right more often.
- Dictation into terminals. Windows Terminal, PowerShell, Command Prompt and Git Bash get the text pasted at the prompt. Line breaks become spaces, so dictated text never runs a command on its own.
- Say "press enter" to send. End a dictation with "press enter" and VoiceWave presses Enter after the text lands. Esc still cancels the Enter.
- Try again after a failed transcription. The pill offers Try again, which transcribes the same recording. The audio stays in memory only, until your next dictation.
- Your first word is no longer clipped. The mic stays ready for 2 minutes after a dictation, and Settings → Audio → Keep mic ready keeps it on while VoiceWave runs.
- Start with Windows (Settings → General, off by default) loads the speech model at sign-in, so the first dictation after a reboot is fast.
- Choose your own dictation shortcut in Settings → Dictation.
- Report a problem and Copy diagnostics in Help and Settings → Diagnostics fill in a GitHub issue with app, Windows, speech engine and microphone details plus recent warnings. Never audio or anything you dictated.
- Log files. A week of daily logs, including any crash. Logs never contain what you dictated.
- More reliability counts in the anonymous usage count, if it's on: setup steps, failed dictation and download types, speech helper crashes, and whether speech runs on the GPU or CPU. Fixed categories only, never content. Settings → Diagnostics turns it all off.
Changed
- Smaller installer, and upgrades remove the old speech runtime first.
- Everything is free, so the Pro screens are gone. The sign-in screen is hidden too.
- Hands-free waits while you think: it ends when you press the shortcut again, or after 30 seconds of silence or 5 minutes.
- Apps running as administrator get the text on the clipboard, and VoiceWave says so.
- Clearer failure messages, and the fonts now ship with the app instead of loading from Google Fonts.
Fixed
- The speech and AI polish helpers now close with VoiceWave, even after a crash.
- Esc reliably cancels a dictation, including a pending "press enter", even when a new dictation starts right after.

Dictation on your graphics card, better accuracy at the start and end of what you say, and a smoother first run.

### New

• **Dictation runs on your graphics card.** On PCs with a dedicated graphics card, VoiceWave now transcribes on the GPU through Vulkan, using the driver you already have; there is nothing extra to download. The Models page shows where dictation runs ("Running on …"), and the new **GPU acceleration** setting can be set to Auto or Off.
• **One-time graphics card check.** The first time the GPU is used, VoiceWave runs a short check (about a minute) before relying on it. The check runs again only after a graphics driver or speech-engine update. Onboarding shows its progress, and a dictation made during the check waits for it instead of running slowly on the processor.
• Existing users with a suitable graphics card get a one-time offer to switch to the GPU model.

### Fixed

• **First and last words are no longer clipped.** The recording kept too little audio before the first word, so short words at the start were misheard ("Restart" became "Please start"). It now keeps half a second on each side.
• Dictionary words no longer force their capitalisation on ordinary words："Things" stays "things" mid-sentence. Names and terms with inner capitals or punctuation (VoiceWave, Claude Code, T-Rex) are still corrected.
• A stuck graphics card is detected within seconds, and the dictation is retried once on a fresh GPU worker. After repeated failures VoiceWave uses the processor for five minutes, then tries the GPU again.
• Dictating with no speech model installed now says so right away.
• Model downloads can be cancelled one at a time, and onboarding can always be finished.
• Complete model files already on disk are picked up at launch instead of downloaded again.

### Changed

• If the graphics card can't be used and the active model is slow on your processor, the home screen offers Small for processors (465 MB). It takes over on the processor automatically, and your model returns to the GPU when it can.
• The optional CUDA add-on has been removed; GPU dictation no longer needs it.

First-run reliability, faster first dictation, and safer storage.

### Fixed

• **Model download progress no longer sits at 0%.** The speech-model download
now grows the progress bar steadily instead of jumping from almost nothing
to done at the very end, and Cancel keeps what was fetched so Install
genuinely resumes.
• **No hidden background download on first launch.** VoiceWave no longer
starts downloading the default model before you choose one, and it no
longer contacts Hugging Face on every launch.
• **First dictation after launch is much faster.** Startup could load the
speech model twice, which made the first dictation take 10–20 seconds. It
now loads once, and if the model is still loading when you finish speaking,
the pill says so.
• Dictating before the model has finished downloading now explains that the
model isn't ready yet, instead of waiting about 45 seconds and failing.
• If text could not be inserted or copied, the pill now shows an error
instead of "Inserted".
• The pill shows Inserted as soon as the text lands, even if saving History
is slow.
• Focused-app detection uses the process name, so window titles such as
"Knowledge base" are no longer mistaken for Microsoft Edge.
• Emoji are typed as a single unit and can no longer be split by other input.
• A briefly locked encryption key file (for example during an antivirus scan)
can no longer cause your history, dictionary, or snippets to be reset. An
existing key is never overwritten.
• History now keeps up to 1,000 recent dictations under every retention
setting (previously 200), and the Settings page says so. Starred entries
export with a ★.

### Changed

• **Model recommendation per PC.** Setup and the Models page recommend the
speech model that suits your hardware and note when Large Turbo will be
slow on this PC.
• Installed models are checked for completeness (all model files present).
The previous "checksum verified" message was inaccurate and has been
removed.
• Lower memory use：the AI polish worker reserves about 450 MB less memory,
and the speech worker no longer loads unused video codecs.

• **Anonymous usage count is now on by default for new installs.** First-run
setup shows what is counted, with a switch to turn it off before the first
dictation. Existing users are asked once (Keep on / Turn off); a stored
"off" is never flipped without that choice, and users who had already
turned it on are not asked again.
• What is sent is unchanged：a random installation identifier, app version,
GitHub distribution channel, one activation marker, and at most one active
marker per day. Audio, transcripts, account identity, filenames, settings,
and device details are never sent. It can be turned off anytime in
Settings → Diagnostics.

Existing installations will receive this release through VoiceWave's signed Tauri updater after the release build finishes.