Parlotype 0.6.0

Added

• Parlotype now speaks 25 languages. The interface is available in
  Bulgarian, Croatian, Czech, Danish, Dutch, English, Estonian, Finnish,
  French, German, Greek, Hungarian, Italian, Latvian, Lithuanian, Maltese,
  Polish, Portuguese, Romanian, Russian, Slovak, Slovenian, Spanish, Swedish
  and Ukrainian — the same 25 languages Parlotype can already transcribe. It
  follows your Windows language automatically; pick a different one under
  Settings → Interface language, and the whole app switches over without a
  restart.

  These translations have not yet been reviewed by native speakers. If
  something reads oddly in your language, please
  open an issue — a real
  correction from someone who speaks it is worth more than anything we can
  check ourselves.

Fixed

• Several settings pages kept the language the app started in. If you
  changed the interface language while Parlotype was running, the Help page's
  hotkey list, the Updates page, and the Cloud providers, Data, llama.cpp and
  Startup pages carried on in the old language until you restarted. They now
  follow the switch like everything else.
• The built-in prompt's name on Settings → Prompts was shown in English
  whatever your interface language. Prompts you create yourself keep the name
  you gave them, untouched.

Under the hood

Translations now go
through a brief/import pipeline — nobody hand-edits a satellite .resx — and
the importer refuses a missing key, a damaged placeholder, or a string left in
English. A structural test fails the build when a settings page composes copy in
C# without refreshing it on a language change, which is what the six stale pages
had in common. Twelve of the new languages independently reported that three
English strings force a grammatical case onto a value substituted at runtime;
that rewording is recorded in the ADR and is still to come
(ADR-069).

Parlotype 0.5.2

Fixed

• The dictation widget now closes itself when you're done. Hold your
  dictation key and it appears as before; release it, and once your words are
  typed in, it fades out on its own about a second and a half later — it no
  longer sits always-on-top over whatever you were dictating into until you
  reach for the mouse and click ✕. Opening it yourself — from the tray, or a
  second launch — still keeps it on screen, and touching a dictation-triggered
  widget (clicking it, dragging it, or opening its language menu) keeps it
  open too.

Under the hood

• The widget's dismissal rule now depends on how it was summoned — a
  dictation gesture auto-hides it, the user opening it directly does not —
  instead of one fixed rule for both
  (ADR-068).

Parlotype 0.5.1

Fixed

• Picking a microphone in Settings now actually changes which one Parlotype
  records from. Every recording used the Windows default input device
  regardless of what was selected under Settings → Microphone — the choice
  was saved but never handed to the recorder. If you'd set a specific
  microphone, this release makes it take effect; if the saved device is no
  longer plugged in, Parlotype falls back to the Windows default as before.

Changed

• The recording waveform reads more clearly. The record button and
  waveform are slightly taller and wider to match the tapered-bar animation
  introduced in v0.5.0.

Parlotype 0.5.0

Highlights

• Parlotype's interface now speaks Russian and Spanish. Choose the language at
  Settings → Interface language, or leave it on System default to follow
  Windows. The change is instant — no restart — and dates and numbers keep
  following your Windows regional settings whatever you pick.
• A translation that can't happen is now labelled, not faked. Turning on
  translate-to-English with a Whisper model that doesn't support it (the
  English-only models and Large v3 Turbo) used to leave the language page and the
  dictation widget showing a normal translation that then silently didn't occur.
  Every surface now shows an amber translation paused instead, points at the
  model as the reason, and leaves the toggle on — pick a multilingual model and
  translation resumes on its own.
• The settings menu scrolls. With Whisper selected the settings list is long
  enough that Data and Help dropped off the bottom of the window, with
  nothing on screen to suggest they were there. The list now scrolls like any
  other.

Added

• Settings → Interface language. System default, English, Русский or
  Español. The settings pages, the dictation widget, dialogs and menus all follow
  the choice, live.

Changed

• The llama.cpp manual-server folder box no longer steers you into a folder that
  gets wiped. For a hand-downloaded llama-server build (Gemma 4 → Manual
  install), the box used to suggest %LOCALAPPDATA%\parlotype\llama-server —
  which a re-run of the installer, and uninstalling, both erase. It now starts
  empty with instructions, and warns you in amber if the folder you point it at
  sits inside Parlotype's install area. A folder you already saved is flagged,
  never silently rewritten.

Fixed

• Every settings page is reachable again. The settings navigation list didn't
  scroll, so on taller layouts — Whisper selected, at the default window size —
  the last entries were clipped off the bottom with no scrollbar and no wheel
  response.
• The language page and dictation widget no longer promise a translation the
  engine won't perform when the chosen Whisper model can't translate. The
  connector, the summary sentence and the widget strip now all read "paused", in
  amber rather than the blue of a live translation, and the toggle stays operable
  so switching back to a capable model just works.

Under the hood

Full rationale for each item is in the linked decision record.

• The interface-language foundation: externalised string resources, a
  live-switching compiled-binding layer, typed format helpers, and a parity check
  wired into both dotnet test and the release gate, shipping with Russian and
  Spanish resource files
  (ADR-064).
• Model-blocked translation is modelled as a distinct, reversible "paused" state
  derived from a single predicate and rendered on every surface, instead of a
  forced-off preference
  (ADR-061).
• Non-installed builds now bind their lifetime to the process that launched them,
  so a killed dotnet run, an IDE stop, or a XAML-previewer refresh can't leave a
  headless copy running that still answers the dictation hotkey
  (ADR-062,
  ADR-063).
• Pack-folder detection for user-supplied paths, plus removal of the vestigial
  LlamaServerDirectory that named a folder nothing ever created
  (ADR-065,
  ADR-066).
• The settings nav list gets a real viewport from a DockPanel, with edge fades
  and deep-link scroll-into-view for a reused window
  (ADR-067).

Parlotype 0.4.4

Highlights

• Push-to-talk holds are no longer cut mid-sentence. Pausing briefly while
  holding your dictation key used to end the recording early — corrupting the
  words either side of the cut and deciding punctuation on half a sentence.
  Now only releasing the key ends it, so a full sentence goes in as one
  transcription, however long you pause mid-way through it.

Changed

• Long recordings get more context before they're capped, and the cap now
  depends on the speech engine instead of a flat 30 seconds: up to 60s on the
  default Parakeet engine (the point past which it starts silently dropping
  words) and 300s on Whisper. Go past it and Parlotype now splits at your next
  pause instead of cutting wherever the cap landed.
• Settings → Speech → Silence timeout now explains itself correctly: it
  only governs toggle-mode dictation. Push-to-talk holds ignore it — releasing
  the key already marks the end of your sentence — and the description used
  to claim otherwise.
• Text now appears only once you release the push-to-talk key, instead of
  trickling in during pauses, since the whole hold is transcribed together.
  A typical sentence still finishes well under a second on Parakeet.

Fixed

• A recording that ran past the old 30-second cap used to be transcribed from
  raw, un-filtered audio; it now goes through the same silence-trimming as
  every other recording.

Under the hood

Full rationale for each item is in the linked decision record.

• Hold-scoped push-to-talk gestures now select a SingleUtterance pipeline
  mode carrying no silence-based cutoff, derived from the gesture rather than
  a setting; per-engine ceilings and boundary-aware splitting on overflow
  (ADR-060).

Parlotype 0.4.3

Highlights

• Parlotype now starts with Windows, so your hotkey works right after a
  reboot. Parlotype only listens for its dictation hotkey while it's running,
  and until now nothing started it after sign-in — a restart left the hotkey
  silently dead until you opened the app yourself. This is on by default; turn
  it off anytime at Settings → Application → Startup.

Added

• Settings → Application → Startup. A single toggle for launching at
  sign-in, with status text that reflects what Windows will actually do — for
  example if you've separately switched Parlotype off in Task Manager's
  Startup apps tab, the page tells you that instead of showing a switch that
  claims to be on while nothing launches.

Under the hood

Full rationale for each item is in the linked decision record.

• Registers via the per-user HKCU\...\Run key (no elevation, no service, no
  scheduled task) pointed at the Velopack install stub, reconciled against both
  the stored preference and Windows' own Task Manager veto
  (ADR-059).

Parlotype 0.4.2

Highlights

• The installer is digitally signed. Windows SmartScreen no longer shows the
  full-screen "Windows protected your PC" warning on install, with the run
  button hidden behind "More info" — this is the first signed release.
• Canceling a dictation while holding Ctrl or Alt now always works. A
  shortcut typed at normal speed mid-recording — a slow Ctrl+C, say — used to
  still transcribe what you'd said and type it into whatever you were working
  in. Now any keystroke during the hold cancels, however fast you type it.

Changed

• Every file Parlotype ships is signed, not just Setup.exe. Update.exe
  and the bundled Whisper/Parakeet/Vulkan native libraries are covered too, so
  Smart App Control and enterprise publisher-rule policies can run Parlotype
  without extra exceptions.

Fixed

• Canceling a dictation while holding Ctrl or Alt now always cancels.
  Previously only a shortcut typed within 300ms of the key-down counted as a
  cancel; anything slower let the recording finish normally and transcribe
  into your target app. Holding Shift is unaffected — typing while Shift is
  held still composes text as before, since dictation and Shift-modified
  typing were never distinguishable.
• Canceling a dictation now actually throws the recording away, instead of
  quietly transcribing it in the background for up to 30 seconds afterward. A
  cancel followed immediately by a new recording could previously be dropped
  because of this.

Under the hood

Full rationale for each item is in the linked decision record.

• Installer and native-library signing via Azure Artifact Signing, invoked
  from inside vpk rather than as a separate post-pack step, over OIDC with
  no stored secret
  (ADR-058).
• Ctrl/Alt holds drop the abort grace window entirely, and cancel now drives a
  real discard path (IAudioPipeline.CancelAsync) instead of draining through
  the normal stop
  (ADR-057).

Parlotype 0.4.1

Highlights

• A guided tour greets you on first launch. It opens the real Transcribe
  and Settings windows and points at the real controls, naming your actual
  configured hotkeys instead of showing generic defaults — so it can't go
  stale the way a screenshot walkthrough would. Replay it anytime from
  Settings → Help.
• Running Parlotype twice is safe now. Opening it again — a Start-menu
  tile, a pinned shortcut, "Run now" after install — used to silently start a
  second, competing copy. Now it just brings the existing window forward.

Added

• First-run tour. An 8-step walkthrough covering recording, the speech
  engine, model selection, cloud engines, and the tray icon. It shows once,
  including on your first launch after upgrading from an earlier version.
• Settings → Help. Replay the tour with Open the tour, and see a live
  reference of your current hotkeys.

Fixed

• Launching Parlotype a second time no longer creates two competing
  copies. Each extra process used to load its own model and race the others
  to answer the same hotkey press, so which one (if either) typed your
  dictation was undefined. A second launch now just brings the existing
  recording window to the front, the same as clicking the tray icon.

Under the hood

Full rationale for each item is in the linked decision record.

• Single-instance enforcement via a named mutex acquired in Program.Main,
  with cross-process activation of the existing window
  (ADR-055).
• The onboarding tour's live UI highlighting, deep links into Settings
  sections, and the repo's first externalized-strings layer
  (ADR-056).
• Release notes are now curated from this file instead of falling back to the
  raw squash-merge commit body
  (ADR-054).

Parlotype 0.4.0

Highlights

• Parlotype installs and updates itself. A real installer replaces the
  hand-unzipped folder, and Parlotype now checks for new versions in the
  background and updates in place — no more downloading a fresh zip every time.
• The download is 82 MB instead of 253–385 MB. One build, one file. The
  confusing "full" and "lite" choice is gone, and after the first install
  updates arrive as patches of a few hundred kilobytes.
• Bring your own cloud engine, if you want one. Two optional cloud
  transcription services (any OpenAI-compatible host, including Groq, and xAI
  Grok) sit alongside the on-device engines. They are off by default, need your
  own API key, and show a persistent Cloud badge whenever one is active.

⚠️ Action required if you used an earlier version

Parlotype's data folder moved from %LOCALAPPDATA%\parlotype to
%LOCALAPPDATA%\parlotype-data, because the installer now owns the old
location and wipes it on uninstall. There is no automatic migration. With
Parlotype closed, and before running the installer:

move "%LOCALAPPDATA%\parlotype" "%LOCALAPPDATA%\parlotype-data"

Skipping this is safe — Parlotype starts with default settings and re-downloads
models on demand. You would just lose your settings, saved API keys and the
models you had already downloaded. See
the migration notes
if you install first and need to recover afterwards.

Parlotype is not code-signed yet, so Windows SmartScreen warns on first run.

Added

• Cloud engines (opt-in). OpenAI-compatible (configurable base URL, so
  OpenAI, Groq or a self-hosted server all work) and xAI Grok. Keys are stored
  encrypted with Windows DPAPI, never in settings.json. Cloud engines always
  auto-detect the language, so the language controls hide when one is selected.
• Hotkeys you can actually choose. Dictation now takes a list of gestures
  instead of one chord, including hold-a-modifier and double-tap-a-modifier.
  New defaults: hold Right Ctrl to talk, double-tap Ctrl to toggle, and
  Ctrl+Alt+Space. The old Ctrl+Shift+Space default is retired — it was
  Parameter Info in Visual Studio and signature help in VS Code. If you had
  picked your own hotkey, it is kept as your only binding.
• Press Escape to throw a dictation away while it is running, instead of
  waiting for it to transcribe into whatever window you were in.
• Settings → Updates. Automatic update checks (on by default, one anonymous
  request to the public GitHub release feed — you can turn it off), when it last
  checked, a manual Check now, and Restart to update once one is staged.
• Settings → Application → Data. See where your data lives, copy the path or
  open the folder, see how much disk your downloaded models use, delete them,
  and opt in to having everything removed if you ever uninstall (off by default).
• "How prompts work" panel on the Gemma 4 prompt settings, explaining the
  {speech_lang} and {text_lang} placeholders and when translation kicks in.

Changed

• GPU acceleration for Whisper is Vulkan-only. The CUDA option is gone. It
  never shipped the NVIDIA libraries it needed, so it only worked if you had
  separately installed the ~3 GB CUDA toolkit — which is why half the runtime
  settings page was toolkit instructions. Measured on LibriSpeech test-other,
  Vulkan is within 8–26% of CUDA on speed with identical accuracy on Small and
  Medium, and better accuracy on Large v3 Turbo. NVIDIA cards are still
  accelerated, through Vulkan. If you had selected CUDA, you are moved to
  Auto, and NVIDIA users on the larger models get roughly 800 MB of RAM
  back. Gemma 4 is unaffected and can still use CUDA builds of llama-server.
• Recording allocates ~10× less memory — about 3 MB/s instead of ~30 MB/s
  during dictation — so long sessions put much less pressure on the garbage
  collector. Voice detection also moved off the microphone callback thread,
  where a slow moment could silently drop audio.
• Dictation never reaches your clipboard history. Injected text is now
  marked to stay out of Win+V history and Cloud Clipboard sync.
• Transcripts are never written to logs — only their length. Log files are
  capped at Information level so debug detail cannot persist.
• Every model download is checked against a SHA-256 digest and fails without
  touching your model folder if it does not match.
• Cloud provider URLs must be HTTPS unless they point at your own machine.
• Settings and saved keys are written atomically, so a crash mid-save can no
  longer corrupt them into a silent reset.
• Parlotype ships with its own waveform icon instead of a placeholder, and the
  executable is now Parlotype.exe.

Fixed

• Switching the Whisper runtime mid-session no longer breaks recording.
  Whisper picks its runtime once per process, so a change only takes effect
  after a restart — but nothing said so, and every record press afterwards
  failed with an error implying broken drivers while leaking a full model's
  worth of RAM per attempt (up to ~3 GB, ~18 GB with large-v3). Parlotype now
  tells you a restart is required, on the settings page and before anything
  loads.
• Push-to-talk no longer misses your key release while a model is still
  loading, and cancelling during a cold load takes effect immediately instead of
  seconds later.
• Removing every hotkey binding sticks. Clearing the list used to silently
  hand the defaults back on the next launch.
• Escape only cancels when pressed on its own. It used to fire with any
  modifiers held, so a Ctrl+Escape binding discarded your dictation, and
  Ctrl+Esc/Alt+Esc were swallowed from the rest of Windows while recording.
• Gemma 4 prompts substitute {text_lang} correctly in custom prompts,
  including while translating, where it previously leaked through as raw text.
  All three built-in prompt bodies now ask for punctuation explicitly.

Under the hood

Full rationale for each item is in the linked decision record.

• Packaging moved to Velopack 1.2.0 with delta updates
  (ADR-053).
• Published output shrank 731 MB → 180 MB by dropping the CUDA runtime
  (ADR-049),
  the never-loaded ONNX Runtime GPU execution providers
  (ADR-050),
  Whisper natives built for other platforms
  (ADR-051)
  and native Skia/HarfBuzz PDBs
  (ADR-052).
• Audio pipeline reworked into channel-joined stages with pooled buffers
  (ADR-045),
  measured by a new micro-benchmark project
  (ADR-044),
  plus a full security audit and its remediations
  (ADR-046).
• Hotkey recognition moved into Core as pure timestamp-driven state machines
  (ADR-047);
  the Whisper runtime latch became a Core contract
  (ADR-048).
• Migrated off obsolete Avalonia AXAML APIs; the benchmark workflow is pinned to
  Windows runners because Whisper.net's native library crashes the .NET test
  host on Linux.

v0.3.0

Highlights

• A new default engine: NVIDIA Parakeet TDT 0.6B v3. It runs on the CPU, is
  substantially faster than Whisper, and detects all 25 supported European
  languages automatically — so there is nothing to configure. Whisper and
  Gemma 4 are still there if you need ~99 languages or translation.

Added

• Parakeet is downloaded automatically the first time you use it (~670 MB, via a
  dialog you can cancel).
• A full-precision Parakeet variant (~2.6 GB) for noticeably better accuracy at
  roughly twice the decode time and three times the RAM. Selectable in settings.

Changed

• The dictation window is a compact, frameless widget (172×112). Drag it by
  the strip along the top, and ✕ or Escape hides it to the tray. Its position is
  remembered between runs, with a fallback if the screen it was on is gone.
• Language controls are hidden for Parakeet, which has no language choice to
  make. Your Whisper and Gemma 4 language preferences survive switching engines.

Fixed

• Releasing a push-to-talk key while the model was still loading no longer
  leaves the recording stuck.