--- title: "Changelog — Voicetypr" description: "Follow every Voicetypr release with dated features, fixes, breaking changes, download notes, and links to the public desktop-app changelog and releases." image: "https://voicetypr.com/voicetypr-og.png" canonical_url: "https://voicetypr.com/changelog" md_url: "https://voicetypr.com/changelog.md" language: "en" --- Product history # Changelog Every release, latest first. Curated from the desktop app’s [CHANGELOG.md](https://github.com/ideaplexa/voicetypr/blob/main/CHANGELOG.md). Plain language, no commit hashes. ## 2.0.5 2026-08-02Latest ### Highlights - More reliable Windows transcription with smarter discrete-GPU selection and safer CPU fallback - Paid licenses remain available through temporary validation outages - Tray recovery, file-upload timelines, and problem reporting are more resilient ### Features - Updates: adds Stable and Beta release channels so you can choose early builds or stay on production releases - Audio feedback: adds optional cues for recording ready, transcription processing, and successful text delivery - Support: replaces the compact report modal with a dedicated page that gathers diagnostics, previews system details, and keeps email and copy fallbacks - Formatting: explains how speech models infer punctuation and shows whether optional AI formatting ran or was skipped ### Fixes - Windows: prefers a discrete GPU on hybrid systems, guards CPU fallback against memory exhaustion, improves transcript spacing, and reduces trailing-silence hallucinations - Hotkeys: accepts injected input so Stream Deck and other external tools can trigger recording - Tray: recovers from temporary startup failures, keeps the dashboard reachable, and includes tray diagnostics in support reports - File uploads: constrains long speaker timelines so Copy and Save remain reachable - Licensing: retries temporary validation failures, repairs stale device activations, and preserves paid access while revalidation is unavailable ## 2.0.4 2026-07-01 ### Highlights - Cloud dictation starts faster — the first request skips the connection handshake - Back-to-back recordings reuse a single connection instead of rebuilding it each time ### Features - Cloud STT: warms the active cloud speech-to-text connection at recording start so the first transcription skips the DNS, TCP, and TLS handshake - AI formatting: pre-warms the active AI-polish connection whenever formatting is enabled with a configured provider - Performance: reuses one pooled HTTP client across requests instead of rebuilding it per call, speeding up back-to-back dictations ## 2.0.3 2026-06-30 ### Highlights - Deepgram keys now validate correctly and report the real problem - Each cloud provider card shows which model it's using ### Fixes - Cloud STT: validates the Deepgram API key against its documented token endpoint and surfaces a precise "model unavailable — check the key's scopes and your plan" message instead of a misleading "invalid key" - Cloud STT: updates Soniox to its current async transcription model - Cloud STT: shows the underlying transcription model on each cloud provider card ## 2.0.2 2026-06-30 ### Highlights - Hotkey engine rebuilt on one native trigger system for combos, holds, and taps - Windows tray no longer crashes on startup and adapts its icon to light or dark ### Features - Hotkeys: unifies key combos, bare-modifier holds, and isolated taps onto one native trigger engine, retiring the old global-shortcut plugin - Hotkeys: lone-modifier recording hotkeys now fire even while a Voicetypr text field is focused, via an in-app fallback - Tray: theme-adaptive tray icon that matches the Windows light or dark taskbar ### Fixes - Tray: never crashes on startup when the system tray fails to initialize on Windows - Windows: ships debug symbols (.pdb) so crash reports symbolicate to real frames - Hotkeys: arms an upgraded combo at startup and no longer swallows bare modifiers on Windows - Hotkeys: clears the old combo hotkey before saving a bare-modifier primary so recording can't double-fire - Windows: guards monitor queries against a stale-handle crash when the display configuration changes ## 2.0.1 2026-06-25 ### Highlights - Crash and reliability hardening across Windows recording, the keyboard hook, and local Whisper - Diagnostics reporting now defaults to opt-out with clearer consent wording ### Features - Diagnostics: crash and error reporting defaults to opt-out with clearer consent wording ### Fixes - Recording: prevents a Windows crash from an unbounded audio buffer size and isolates device-callback panics so a hiccup can't abort the app - Recording: finalizes recorder teardown on a path independent of device disconnects - Hotkeys: wraps the Windows low-level keyboard hook so a hook failure can't take down input - Transcription: hard-timeouts in-process Whisper decoding to stop indefinite "Transcribing…" hangs, and aborts the watchdog before remapping completed results - Hotkeys: single-tap triggers now work in onboarding, matching the Shortcuts page - Settings: avoids a window-state panic when changing the pill offset - Models: keeps a known model selected on startup, opens native microphone settings, and shows animated progress for Parakeet downloads - Onboarding: exposes the model Repair action for parity with the dashboard - Permissions: routes grant buttons to the native settings panes ## 2.0.0 2026-06-25 ### Highlights - Cloud speech-to-text arrives: Soniox, OpenAI, Groq, Deepgram, and Cohere, alongside local models - AI polish rebuilt natively in Rust, with a unified transcription engine spanning local, cloud, and remote - A new CLI, speaker diarization, richer history, and Windows GPU/CPU acceleration ### Features - Speech-to-text: adds cloud transcription providers — Soniox, OpenAI, Groq, Deepgram, and Cohere — selectable per recording - AI formatting: rebuilds AI polish natively in Rust for OpenAI, Anthropic, Gemini, and custom OpenAI-compatible endpoints, with a searchable model catalog and a deterministic fallback when formatting fails - Transcription: unifies local, cloud, and remote transcription behind one engine, with cancel-anywhere, a decode watchdog, silence handling, and device-disconnect recovery - Upload: adds cloud speaker diarization for file uploads (Deepgram, Soniox), producing speaker-attributed transcripts you can save as .txt or .md - CLI: installs the `voicetypr` command on your PATH so terminals, scripts, and AI agents can drive transcription directly, with consistent --json output - History: stores rich per-entry metadata (source, engine, language, duration, diarized) with source, app, and date filters, plus a Show original / Show formatted toggle - Windows: choose GPU, CPU, or Auto transcription acceleration in Settings and onboarding - Windows: adds a Microsoft Store (MSIX) distribution option, with in-app updates managed by the Store for Store installs - Shortcuts: streamlines the recording shortcut to one flow, with first-class single-key push-to-talk, a bindable Toggle AI formatting key, and lone-modifier hold-to-talk ### Fixes - Transcription: shows short, plain messages in the recording overlay instead of long internal error strings, pointing auth and model failures to Settings - AI: validates custom OpenAI-compatible connections the way formatting actually runs, so Test no longer falsely passes on a public models list ## 1.13.0 2026-06-15 ### Highlights - A quiet pause no longer throws away your recording — silence now shows a non-destructive warning - Bug reports include full system specs (GPU, CPU, model) for faster support - More reliable recovery from stuck or auto-stopped recordings ### Features - Recording: replaces the destructive silence auto-stop with non-destructive warnings, so a quiet pause no longer discards in-progress audio - Bug reports: collect full system specs (GPU, CPU, model) so issues are faster to diagnose ### Fixes - Recording: recovers from autonomous recorder stops that previously left recording stuck - Recording: runs stop cleanup on the recorder-error path so a failed recorder no longer hangs the session - Recording: requires sustained voice before treating audio as speech, reducing false starts from brief noise - Windows: pauses media players that under-report pause support while recording - Bug reports: skip software display adapters during GPU detection so the real GPU is reported ## 1.12.10 2026-06-10 ### Highlights - Local Whisper transcription no longer gets stuck in Transcribing on slow CPU-only systems - Short CPU dictation is faster by avoiding hidden Whisper retry passes - No-speech and timeout cases now give clearer feedback instead of silently inserting nothing ### Fixes - Transcription: disables Whisper temperature fallback retries for the CPU profile to reduce latency on short recordings - Transcription: adds bounded timeout and cooperative cancellation handling for local Whisper recording transcription - Feedback: treats empty output, [SOUND], and [BLANK_AUDIO] as no-speech cases with a clear message - Release: pins Windows CI and release runners to VS 2022 so Windows builds do not fail on unsupported VS 2026 CMake generator detection ## 1.12.9 2026-06-08 ### Highlights - Model downloads now verify against current upstream checksums - Windows toggle hotkeys no longer start and immediately stop from repeated key events - Onboarding shows clearer download errors and only continues once a model is ready ### Fixes - Models: replaces stale checksum metadata so valid Whisper downloads are no longer deleted after verification - Onboarding: surfaces terminal download errors and aligns model readiness with the main Models tab - Windows: ignores duplicate toggle hotkey press events while the shortcut is held, preventing empty recordings ## 1.12.8 2026-06-05 ### Highlights - Model management stays responsive during downloads - Settings can refresh model status without interrupting an active download ### Fixes - Models: keeps the UI responsive while a download is in progress - Downloads: prevents status refreshes from racing active model operations ## 1.12.7 2026-06-05 ### Highlights - Windows model downloads and catalog metadata repaired - Windows Vulkan sidecar warm-up is safer and less noisy ### Fixes - Windows: repairs local model catalog downloads and related runtime packaging - Windows: warms the optional Vulkan sidecar more reliably without making the main app depend on Vulkan - Updater: backs out a passive installer path that could hide elevation prompts on some machines ## 1.12.6 2026-06-04 ### Highlights - Windows transcription fallback is faster and safer on machines without GPU acceleration ### Fixes - Windows: optimizes transcription fallback so CPU-only recovery is more usable when GPU acceleration is unavailable ## 1.12.5 2026-06-02 ### Highlights - Windows GPU transcription hotfix for the 1.12.4 Vulkan sidecar regression - Successful Vulkan sidecar responses now parse correctly instead of falling back to slow CPU mode ### Fixes - Windows: fixes Vulkan sidecar timing fields so GPU responses parse correctly - Transcription: prevents the 1.12.4 GPU path from falling back to unusably slow CPU transcription after successful sidecar work ## 1.12.4 2026-06-01 ### Highlights - Windows now launches safely on machines without Vulkan drivers - Updates ask before downloading or installing unless automatic installs are enabled - Hotkey and push-to-talk reliability improvements with better support diagnostics ### Features - AI: restored the native Anthropic provider with current model aliases ### Fixes - Windows: moved optional Vulkan Whisper acceleration into a sidecar so the main app remains CPU-safe - Updater: shows update availability without silently downloading or installing by default - Windows: fixed updater elevation by using the basic UI installer flow - Recording: prevents duplicate push-to-talk press events from starting overlapping flows - Support: adds hotkey diagnostics/report context for shortcut failures - macOS: keeps Parakeet sidecar JSON responses clean when CoreML emits native diagnostics ## 1.12.3 2026-04-29 ### Highlights - Auto-paste toggle so transcripts can land in history without inserting into the active app - In-app bug and crash reports with redacted log excerpt and system info ### Features - Settings: auto-paste toggle keeps transcripts in history without auto-inserting them - Support: in-app bug and crash report submission with system info and a redacted app-log excerpt ### Fixes - Support: hardened report validation, stale-submit guards, fallback behavior, and log redaction - Settings: restored scrolling on Upload, General, Help, Advanced, and Formatting pages after the sidebar layout change ## 1.12.2 2026-04-27 ### Highlights - Long recordings no longer truncate — captures hold their full length - Sentence spacing preserved on insert; punctuation matches the host app - Windows: WebView2 offline installer bundled for clean first-run installs ### Fixes - Transcription: fixes truncation in long recordings; preserves explicit insertion boundaries - Transcription: preserves sentence spacing on insert and uses compatible punctuation matching - Recording: hardened push-to-talk startup, license fallback, and abort handling - Windows: reduced crash paths and bundled the WebView2 offline installer for clean first-run installs - UI: kept sidebar navigation accessible and routed settings nav to the correct tab - License: shows the Manage License button when a license is already active - AI: persists AI formatting across restart by warming cached credentials from secure storage - Updater: removed stale v1 dialog config and uses a versioned update marker for post-update UX - Autostart: moved autostart toggles to backend-owned commands for atomic OS state management ## 1.12.1 2026-02-22 ### Highlights - GPT-5 token + temperature compatibility for OpenAI users ### Fixes - OpenAI: handles token and temperature compatibility for gpt-5 ### Breaking changes - AI: removed the standalone Anthropic provider — Anthropic models still reachable via OpenAI-compatible config ## 1.12.0 2026-02-21 ### Highlights - Intel Mac (x86_64) support - Multi-provider AI formatting: OpenAI, Anthropic, Gemini, with curated model lists - Pause-media-during-recording so podcasts don't talk over you ### Features - macOS: Intel Mac (x86_64) support - AI: multi-provider AI formatting (OpenAI, Anthropic, Gemini) with curated model lists - AI: language-aware formatting and stale microphone validation - Recording: sound-on-end with toggle, plus a pill indicator mode dropdown (Never / Always / When Recording) - Recording: indicator position setting with 6 options and configurable edge offset - Audio: pause media during recording so background audio doesn't talk over your captures - Crash reporter UI with GitHub issue integration ### Fixes - Windows: ARM64 crash fix — Vulkan GPU disabled, CPU threads optimized - Windows: bundled runtime, hardened fresh-install behavior, hides console flash on recording start - Audio: prevents app hang when an audio device doesn't respond to cleanup; platform-specific stream cleanup - Audio: delay after start sound for Bluetooth headset compatibility - Media: improved pause/resume reliability; uses osascript playback toggle on macOS - AI: isolates custom and OpenAI provider key/config handling - UI: removed unwanted white ring border; consistent ring styling on pill and toast - UI: slider now shows a single value instead of a range - Networking: per-request HTTP clients; media resumes on recording errors ## 1.11.2 2025-12-16 ### Highlights - Session-aware auto-updates with notification support - macOS: dock icon now follows main window visibility ### Features - Updater: session-aware auto-updates with notification support - macOS: dock icon toggles based on main window visibility ### Fixes - Updater: auto-installs updates in the background; About section uses the unified update service - Windows signing: avoids dropping empty password; uses pnpm tauri signer; correct key file path flag ## 1.11.1 2025-12-14 ### Highlights - Pill UI simplified to a 3-dot indicator with a separate toast window - Critical events queue for reliable delivery - Clipboard retention preference ### Features - Pill UI simplified to a minimal 3-dot indicator with a separate toast window - Pill/toast moves lower on screen and auto-repositions on monitor change - Critical pill events queue for reliable delivery - Clipboard retention preference - AI model availability and auto-selection improvements ### Fixes - Microphone permission prompts deferred until onboarding completes - Audio device listeners clean up properly - AI config auto-selection improvements - OpenAI config UX: Update button, better switch visibility - Recording feedback: sound, throttle, and duration gates - Toast notification: dynamic width and clearer error messages - Recording errors routed via pill toast - Microphone list synced with system - Watchdog timeout treated as the only timeout signal - Windows: device ID fallback and license status timeout - Logs: API keys redacted - License: error and reset edge cases hardened ## 1.10.0 2025-10-24 ### Highlights - Parakeet MLX sidecar integration for Apple Silicon - Parakeet v2/v3 with version-aware download/load/delete - Tray model selection unified with the dashboard ### Features - Parakeet MLX sidecar integration (macOS Apple Silicon) - Parakeet v2/v3 version-aware download / load / delete with FluidAudio 0.6.1 - Tray: unified model selection that stays in sync with the dashboard - Replaced Python Parakeet with a Swift / FluidAudio implementation - Windows: updated bundled ffmpeg ### Fixes - Audio: prevents empty / ultra-short recordings from erroring; mode-specific min durations (PTT ≥ 1s, Toggle ≥ 3s); auto-recovers Error → Idle - macOS: bundles ffmpeg/ffprobe via platform config so packaged-app normalization works - Parakeet: model selection persists across restarts - Parakeet: scoped to macOS builds; sidecar build issues resolved - Tray: gates selection until onboarding completes Earlier releases live on GitHub, including the unedited commit history. [View on GitHub →](https://github.com/ideaplexa/voicetypr/releases)