v5.0.0

VoiceToClipboard v5.0.0 — Release Notes (DRAFT)

Highlights

• Idle transparency now ON by default (still disableable in Settings).
• Settings window no longer minimizes or gets buried when changing options during fullscreen games.
• Silence-threshold tools (slider, Reset, Auto-Calibrate, live mic meter) now always available — even with Auto-Stop off.
• Live Mic Level meter runs whenever Settings is open (previously only during calibration).

Fixes

• Settings UI: fixed WIDGET_STYLES ReferenceError that could kill the settings panel init.
• Settings window: removed child-window linkage to the widget; pinned to topmost ('screen-saver' level); re-pinned on reopen/show.
• Always-on-top: settings saves no longer silently downgrade the widget's topmost level.
• Config load resilience: settings UI falls back to last-known-good values if the first config read fails/stalls.
• Dropdown self-heal + external activation handling (markSettingsExternallyActivated) for reliable remote refresh.
• Visualizer starts instantly on record instead of waiting for first audio frames.
• Microphone stream leak/race fixes on rapid start/stop.
• Gemini engine: request timeout + response sanitizer hardening.
• ~40 tree-wide robustness fixes from the pre-release audit (phases 1-2).

Notes

• Recordings folder only grows when "Save recordings" is enabled (default OFF).
• Known cosmetic: CSP inline-style warnings in app.log are harmless.

VoiceToClipboard v4.2.0

VoiceToClipboard v4.2.0

Highlights & Enhancements

• Precision Theme Contour Geometry: Recording button contour, tracer rings, aura blooms, and expanding sonar shockwaves now dynamically match the geometric shape of each widget theme (Crimson circle, Ocean pill, Aurora fluid blob, and Neon Terminal square) with zero clipping or rotational deformation.
• FastConformer Transducer as Recommended Default: NVIDIA NeMo FastConformer Transducer INT8 (mini-multilingual, ~102 MB) is now the default recommended model tier, delivering ultra-fast, high-accuracy offline transcription across English, Spanish, German, French, and 10 languages.
• Enhanced Transcribing Feedback: High-visibility dual-arc neon spinner with smooth, in-place rotation next to the TRANSCRIBING badge status across all themes.
• Settings & Pipeline Fixes:
  • Model download progress spinner with persistent DOM tracking.
  • Preserved engine and output settings during model downloads.
  • Multi-key API cycling support without array truncation.
  • Keyboard accessibility enhancements (Tab navigation guard & Enter/Space engine toggling).
  • Sound preference checks for complete silent operation.

Downloads

• VoiceToClipboard-4.2.0-Setup.exe: Full Windows installer with auto-update support.
• VoiceToClipboard-4.2.0-portable.exe: Standalone portable executable.

VoiceToClipboard v4.1.1

Working build — no known bugs.

VoiceToClipboard 3.1.0

VoiceToClipboard 3.1.0

Four fixes you asked for — performance, feedback, clarity.

⚡ Model unload is now instant

• Power-Saving Mode unloads the model right after transcription and frees the RAM immediately: sherpa now calls any explicit native release handle and triggers V8 garbage collection; Electron starts with --expose-gc. The ~0.3–1.2 GB of model memory no longer lingers.
• Vosk unload also triggers GC for consistency.

🔄 Model status updates itself

• After downloading (or removing) a model, the settings and main-widget model cards refresh automatically — no more manual click needed.

👀 No more "frozen" downloads

• The extraction/verification phase (after the download bar hits 100%) now shows an animated spinner + pulsing progress bar, so it's obvious the app is working.

🏆 Clearer model recommendations

• The two highest-accuracy models (big, zh-big) get a "Best quality" chip in the dropdown.
• A brief note in the local-models section reminds: RAM is used only while transcribing when Power-Saving Mode is on, and bigger models = higher accuracy.

Misc

• README version badge synced to 3.1.0.

Windows installer (Setup + portable): run npm run build on Windows after pulling — assets are attached to this release when built.

VoiceToClipboard 3.0.0

v3.0.0 — i18n + re-transcribe + click-through + ring fix

Auto-stop reliability (the big one)

• Adaptive noise floor: the silence threshold now tracks your room's ambient noise (threshold +8 over floor), so it works whether you calibrated in a noisy or quiet room — no more recordings that never stop.
• Crash-proof animation loop: an exception anywhere in the visualizer used to silently kill BOTH the ring animation AND the auto-stop timer. The loop is now exception-proof and errors are logged.
• Blip-tolerant timer: keyboard clicks and pops no longer hard-reset the 2s silence countdown — they only eat their own duration.
• Telemetry: every recording logs record start, VAD summary (min/max/avg level, effective threshold, armed flag) and a dead-mic watchdog to app.log — a misbehaving mic is diagnosable in seconds.

New in this build

• Full i18n (EN / ES / ZH): the entire app — settings, statuses, tooltips, tray menu, paste bubble/toast — in your language. Auto-detects Windows language (Spanish on this PC), switchable in Settings → Interface language. All offline, 139 strings, hand-translated.
• Re-transcribe button (↻): after a transcription, a ↻ button appears in the top bar — re-runs the LAST audio with the current engine/model (switch model first, then retry). Audio is kept after success until the next recording.
• Click-through fix: only the mic circle, top-bar buttons and retry button capture clicks now — the invisible margins pass clicks to the app underneath. No more phantom click-eating.
• Ring clipping fixed: visualizer canvas 90→150 px — the music ring around the mic never gets cut off anymore.
• Minimalist paste bubble: dark glass card, green dot, red key pill, 2-line START…END clipped transcript (long speeches no longer make a giant bubble). Toast style also clipped.

Models

• Per-language recommendation hints in Settings (Spanish → mini/whisper-base, Chinese → zh-light).
• Code-switching-aware prompts + SenseVoice language hint passed per UI language.

VoiceToClipboard v2.2.0

VoiceToClipboard v2.2.0

Local offline models are now the default. Gemini API is the optional second choice.

🔧 What's new in v2.2.0

• Offline Models = default engine. The app no longer asks for a Gemini API key out of the box — it uses your installed local model. Gemini API moved to second position for when you want it.
• RAM-based ⭐ recommendation. The model dropdown recommends the tier that matches your PC:
  • ≤ 4 GB RAM → Tiny (~250 MB)
  • ≤ 8 GB RAM → Mini (~270 MB)
  • ≤ 16 GB RAM → Light / Omnilingual (~550 MB)

  • 16 GB RAM → Big / Parakeet (~950 MB)

• 🔔 Sound when transcription finishes. A soft two-tone chime plays when your transcript is copied — toggle it off in Settings → Recording.
• Widget that never feels dead. While your cursor is anywhere over the widget it responds fully (click, drag, focus) — click-through only kicks in when the cursor leaves.
• Stays on top while gaming. Windowed-fullscreen games can't bury it: the moment you hit the hotkey or click record, the widget pops above the game (without stealing game input). Fully event-driven — no background timers, no constant re-asserting.
• New: single-file portable exe (VoiceToClipboard-2.2.0-portable.exe) alongside the portable zip and the NSIS installer.

🎙️ Everything from v2.1.0 (unchanged)

• Six verified offline multilingual models (Tiny / Mini / Chinese+English Light / Light ⭐ / Big / Chinese+English Big) — auto language detection, no language setting ever.
• Auto-Calibrate background-noise threshold (5-second "don't talk" test with live meter).
• Wake-peek settings pill, futuristic model picker with download size + RAM per row.
• Gemini 2.5 Flash cloud transcription available as an option.

14/14 automated tests pass. Verified live: local default, RAM recommendation (32 GB → Big), hover interactivity, and packaged builds.

VoiceToClipboard v2.1.0

VoiceToClipboard v2.1.0

Six offline models, automatic background-noise calibration, and a Settings panel you can always find.

🎙️ Six local offline models (all auto-language — no language setting, ever)

Model	Languages (auto)	Download	RAM (worst-case estimate)
Tiny · FastConformer CTC	EN / DE / ES / FR	~98 MB	~250 MB
Mini · FastConformer Transducer	10 languages (EN/DE/ES/FR/IT/PL/RU/UK/HR/BE)	~102 MB	~270 MB
Chinese + English (Light) · SenseVoice	Mandarin, English, Cantonese, Japanese, Korean	~158 MB	~400 MB
Light · Omnilingual 300M v2 ⭐ Recommended	1600+ languages	~279 MB	~550 MB
Big · Parakeet-TDT 0.6B v3 INT8	~25 European languages	~465 MB	~950 MB
Chinese + English (Big) · FireRedASR2	Mandarin + English (code-switching)	~496 MB	~1.1 GB

• All six were downloaded through the real install pipeline and verified live on Windows x64.
• RAM figures include ~⅓ headroom for the transcription spike (a model that idles at ~400 MB is listed at ~550 MB).
• The Light (Omnilingual) model ships as the default — one model, 1600+ languages, ~389 MB on load.

⚡ Auto-Calibrate (background noise threshold)

• One simple test: don't talk, let the background noise play for 5 seconds (3/5/10 s selectable).
• Live volume meter appears only while the test runs, then auto-hides.
• Threshold is set from the noise percentiles + a comfortable margin, so silence never cuts you off mid-pause.

🔧 Settings you can always find

• Wake-peek pill: the settings/close pill now appears the moment your cursor touches the widget (auto-hides after ~3 s on empty space). No more hunting for Settings.
• Fixed a deadlock that could permanently hide the pill after a drag.
• Futuristic model picker: simple tier names (Tiny / Mini / Light ⭐ Recommended / Big / Chinese + English), each row shows model · download size · RAM.
• Frameless, close-only Settings window with ecoMode (loads the model only while transcribing) ON by default.

🚀 Deployment

• Portable build (VoiceToClipboard-2.1.0-portable.zip) and NSIS installer (VoiceToClipboard-2.1.0-Setup.exe).
• Installer includes latest.yml + blockmap for auto-update.

Gemini cloud transcription (2.5 Flash) remains available alongside the local models.

VoiceToClipboard v1.0.0

Voice-to-clipboard in one click. Tiny always-on-top Windows widget: press Ctrl+Alt+V (or click the mic), speak, and the transcript lands in your clipboard.