Hushpen 2.3.3

Hushpen 2.3.3: quieter, quicker cue sounds with soft ends, and a flow bar that fits its text and moves calmly.

Improved

• The cue sounds are half as loud by default: Sound volume starts at 12.5% instead of 25%. If you never changed it, the new default applies; a volume you picked stays. The slider now moves in 2.5% steps.
• The cue sounds play 1.25x faster, so they are a fifth shorter and a little higher. Mute other audio while dictating starts sooner after the start cue.
• Every cue sound fades in and out, so none clicks at its start or end.
• The flow bar fits its text. Pasted, Copied, Cancelled, and notices no longer sit in wide pills, and the pill keeps its size under the pointer.
• The flow bar moves calmly between states. Its width eases to the new size, the old content fades out before the new one fades in, and the listening bars turn into the processing wave where they stand.

Fixed

• On macOS, a square and a round shape no longer flash as the flow bar hides. The bar fades out in place and has no window shadow.

Downloads

• macOS (Apple silicon, macOS 13 or later): Hushpen-2.3.3-macos-arm64.dmg or .zip. Signed with Developer ID (SURAJ MANDAL, GF57JXJF5A) and notarized by Apple.
• Linux x64 and arm64 (X11): AppImage, deb, rpm, and tar.gz.
• SHA256SUMS lists the checksum of every installer. latest.json is the signed update manifest, so Hushpen 2.3.2 with the update check on offers this version.

Install steps: https://github.com/surajmandalcell/hushpen#install

Hushpen 2.3.2

Hushpen 2.3.2: Recover a cancelled dictation, a balanced flow bar with round ends, and less memory when you switch speech models.

New

• Recover a cancelled dictation. After Esc or the cross, the flow bar shows Cancelled with a Recover button for 4 seconds. Recover transcribes the dictation and pastes it as if you had finished it.

Improved

• The flow bar has truly round ends, also on macOS, and every state is balanced: a round button or mark at both ends or at neither, with the text or the wave centered between them. Its sizes changed a little.
• Pasted with Undo paste leads with a soft green done tick.
• The wave is a little shorter at its ends, so it tapers into the round ends.
• The speech engine runs as its own program, hushpen-asr, so Activity Monitor and ps show it by name next to hushpen-llm and hushpen-tts.

Fixed

• A switch of the speech model uses less memory. Hushpen now frees the old model before it loads the next one, so the two are never in memory at once.

Downloads

• macOS (Apple silicon, macOS 13 or later): Hushpen-2.3.2-macos-arm64.dmg or .zip. Signed with Developer ID (SURAJ MANDAL, GF57JXJF5A) and notarized by Apple.
• Linux x64 and arm64 (X11): AppImage, deb, rpm, and tar.gz.
• SHA256SUMS lists the checksum of every installer. latest.json is the signed update manifest, so Hushpen 2.3.1 with the update check on offers this version.

Install steps: https://github.com/surajmandalcell/hushpen#install

Hushpen 2.3.1

Hushpen 2.3.1: a fix for VoiceOver in the menu bar.

Fixed

• The menu bar item now has a name that VoiceOver reads: "Hushpen".

Downloads

• macOS (Apple silicon, macOS 13 or later): Hushpen-2.3.1-macos-arm64.dmg or .zip. Signed with Developer ID (SURAJ MANDAL, GF57JXJF5A) and notarized by Apple.
• Linux x64 and arm64 (X11): AppImage, deb, rpm, and tar.gz.
• SHA256SUMS lists the checksum of every installer. latest.json is the signed update manifest, so Hushpen 2.3.0 with the update check on offers this version.

Install steps: https://github.com/surajmandalcell/hushpen#install

Hushpen 2.3.0

Hushpen 2.3.0: private, local dictation for macOS and Linux, now with Parakeet as the default speech model, S1-mini for AI cleanup, cleanup profiles, word packs, cloned voices, and Audio enhance.

New

• Parakeet v3 Q4 is the default speech model. It is fast and accurate for English and 24 other European languages. The speech list is now four models: Parakeet v3 Q4, Whisper Large v3 Turbo Q5 for Hindi and other languages, Whisper Small Q5 for translation to English, and Whisper Base Q5 for older or slower computers. The English-only and full-size models are gone. If you used one of them, Hushpen moves you to Parakeet and deletes the old file.
• S1-mini by Superwhisper for AI cleanup. It replaces Gemma 4 E2B: a 484 MB download instead of 3.1 GB, and it answers in a moment. A Gemma choice moves to S1-mini, and the Gemma file is deleted to free its disk space. Command Mode needs Qwen3.5 2B or an endpoint.
• Cleanup profiles. Settings > Cleanup has a Profile pop-up: General, Message, Email, and Notes. The Prompt sheet shows exactly what the cleanup model gets. With Qwen3.5 2B or an endpoint you can edit the prompt and reset it; S1-mini shows its fixed prompt and three choices.
• Word packs in the Dictionary. Indian names, Indian places, Hinglish words, and tech terms. Turn a pack on and its words count like your own. Preview shows the words.
• Mute other audio while dictating. Music and videos go quiet while you speak and come back when you stop, also after a crash. It is on by default.
• When idle. Settings > Engine chooses what happens to a loaded model after a quiet time: keep it in memory, page it out (Linux), or unload it.
• Voices that clone. Read aloud has Chatterbox Turbo and Chatterbox Nano. They read with their own voice or with a voice you record or pick from an audio file. Manage Voices keeps your voices: play, rename, delete, and set the default.
• Mix two Kokoro voices. Mix with and Mix in Read aloud blend a second voice into the first, from 10% to 90%.
• Read aloud in History. Each text read to the end is kept in History with its audio. Read aloud again plays it with the current voice.
• Audio enhance. Settings > Audio makes speech clearer for the transcript. Auto is on by default. Custom has a tuning sheet with low cut, bass, treble, and noise reduction, and Find best settings tests your own reading to pick the values with the fewest wrong words.
• Search in Models and in long menus. Models has a search field (Cmd+F, Ctrl+F on Linux). A pop-up menu with more than eight options, such as Language, opens with a search field.
• A taller window. Drag the bottom edge to make the window up to 900 pixels high. Hushpen opens at the same height next time.
• Follows Reduce Motion. Views, sheets, menus, and switches now move into place. With Reduce Motion on in the system settings, every change shows at once, and a change of the setting applies at once.
• A new app icon: a statue whispering into a fountain pen. The menu bar shows a quill writing a sound wave.

Improved

• Start a new dictation at once, even while the last one is still being transcribed. Texts go in in the order you spoke them.
• Silence is skipped before speech recognition, so long pauses and quiet noise give no made-up words.
• Files is now Manual and sits after History. Its Transcribe and Read aloud control is in the title bar, and Read aloud has its status, Export Audio, and Play in the first row.
• The flow bar comes in faster and looks calmer: a small wave while your words are processed, Pasted for one second, and a round Undo paste button for four seconds. On macOS it shows at once for each new dictation.
• History details use icon buttons, with Copy beside the facts. Re-paste is gone; use Copy.
• The window opens on the view that was open when you quit.
• Every slider shows its value beside the track.
• Input boost now starts at 150% and the sounds at 25%. An older install that still has the old values saved (a 100% boost or 50% sounds) moves to the new ones once.
• Model details show the measured memory and how the weights are stored.
• Kokoro keeps its natural pauses at punctuation again.
• A model whose files changed in this version, such as Kokoro, updates itself in the background. Models shows Updating with the percent instead of Failed verification. If the network stops, Hushpen tries again when you open Models.

Fixed

• A long transcript on Home was cut off at both sides. It wraps again.
• A new recording that started right after a stop could lose the microphone.
• Kokoro could make broken sound on short text.
• The window could stop for about a second while Read aloud saved the audio of a long text.

Downloads

• macOS (Apple silicon, macOS 13 or later): Hushpen-2.3.0-macos-arm64.dmg or .zip. Signed with Developer ID (SURAJ MANDAL, GF57JXJF5A) and notarized by Apple.
• Linux x64 and arm64 (X11): AppImage, deb, rpm, and tar.gz.
• SHA256SUMS lists the checksum of every installer. latest.json is the signed update manifest, so Hushpen 2.2.0 with the update check on offers this version.

Install steps: https://github.com/surajmandalcell/hushpen#install

Hushpen 2.2.0

Hushpen 2.2.0: private, local dictation for macOS and Linux, now with shortcut switches, a Read aloud shortcut, an input boost, and new sounds.

New

• Turn a shortcut off. Each shortcut in Settings has a switch. A shortcut that is off keeps its keys for later. With Hold to dictate off, the hold key still starts hands-free with a double tap or Hold+Space.
• Reset one shortcut. A reset button next to a changed shortcut puts it back to its default. Reset to default still resets all of them.
• Read aloud shortcut. Select text in any app and press Ctrl+Cmd+S (Ctrl+Alt+S on Linux) to hear it with the Read aloud voice. Press it again to stop.
• Input boost. A slider in the Audio section of Settings makes a quiet microphone louder, up to 200%. A change is heard at once, also in the microphone test.

Improved

• The flow bar opens on the screen that has the pointer, so it follows you to a second monitor. A position you drag it to applies on every screen.
• On macOS 26 and later the app icon fills its tile like other apps, with no gray border around it.
• New sounds for the start, the finish, and the cancel of a dictation.

Fixed

• Right after a speech or voice model loaded, its status could still say that no model was loaded.

Downloads

• macOS (Apple silicon, macOS 13 or later): Hushpen-2.2.0-macos-arm64.dmg or .zip. Signed with Developer ID (SURAJ MANDAL, GF57JXJF5A) and notarized by Apple.
• Linux x64 and arm64 (X11): AppImage, deb, rpm, and tar.gz.
• SHA256SUMS lists the checksum of every installer. latest.json is the signed update manifest, so Hushpen 2.1.0 with the update check on offers this version.

Install steps: https://github.com/surajmandalcell/hushpen#install

Hushpen 2.1.0

Hushpen 2.1.0: private, local dictation for macOS and Linux, now with file transcription, Read aloud, and shortcuts with any key.

New

• Files. One workspace for audio files and text. Transcribe: drop audio files anywhere on the window (WAV, MP3, FLAC, Ogg Vorbis, M4A, AIFF, up to 12 hours each) and they run in a queue with the time left. Translate to English, then Copy, Export (TXT, SRT, VTT, JSON), Show in History, or Retry. Long files stream through the engine with flat memory use.
• Read aloud. Type text or open a .txt or .md file, and hear it with the system voice or with the local Kokoro voice. Play, Stop, and Export Audio to WAV. A text file dropped on the window opens here.
• Models. A details sheet for every model: what it is for, languages, speed and accuracy, memory, license, source page, folder, and checksum. More Whisper models, a free-disk check before a download, and a Voices section for Kokoro.
• Settings in one list grouped like System Settings, with a search field (Cmd+F on macOS, Ctrl+F on Linux).
• Shortcuts with any key. Every key the system reports can be a shortcut: media keys such as Play/Pause, the keypad, F1 to F35, Escape with other keys, both Ctrl keys together, and chords such as Cmd+Tab or Ctrl+Q.
• Shortcut conflicts you can fix. A shortcut that macOS or another app already uses is saved, and its field says who uses it ("macOS uses this shortcut for "Move focus to next window"."), with one line of steps and an Open Settings button. It starts to work as soon as it is free.
• Cmd+Tab as a shortcut on macOS. While it is your shortcut, Hushpen turns the app switcher off, and turns it on again when you change the shortcut, quit, or the app ends any other way.
• About has a new header with the version and Check for Updates, and links to the website, GitHub, issues, licenses, and the log folder.
• Linux look. With a compositor the window matches macOS: rounded corners, a see-through sidebar that KWin blurs, and a hairline edge. Without one it stays square and solid.

Improved

• Home groups dictation and the microphone, with a microphone test and level meter.
• Pop-up menus float above the page, so no group cuts them off.
• Focus rings, row edges, and separators line up on every view.
• History names the source of a transcribed file "File".

Downloads

• macOS (Apple silicon, macOS 13 or later): Hushpen-2.1.0-macos-arm64.dmg or .zip. Signed with Developer ID (SURAJ MANDAL, GF57JXJF5A) and notarized by Apple.
• Linux x64 and arm64 (X11): AppImage, deb, rpm, and tar.gz.
• SHA256SUMS lists the checksum of every installer. latest.json is the signed update manifest, so Hushpen 2.0.0 with the update check on offers this version.

Install steps: https://github.com/surajmandalcell/hushpen#install

Hushpen 2.0.0

Hushpen 2.0.0: private, local dictation for macOS and Linux.

• macOS (Apple silicon, macOS 13 or later): Hushpen-2.0.0-macos-arm64.dmg or .zip. Signed with Developer ID (SURAJ MANDAL, GF57JXJF5A) and notarized by Apple.
• Linux x64 and arm64 (X11): AppImage, deb, rpm, and tar.gz.
• SHA256SUMS lists the checksum of every installer. latest.json is the signed update manifest.

Install steps: https://github.com/surajmandalcell/hushpen#install

v1.1.0: Docs Redesign & App Icon Update

Full Changelog: v1.0.2...v1.1.0

ASR Pro 1.0.2

ASR Pro 1.0.2 desktop release.

Platform	Packages
macOS arm64	DMG, ZIP
Windows x64	Setup EXE, Portable EXE
Linux x64	AppImage, deb, rpm, tar.gz

Validation	Result
Unit tests	npm test, 4 files and 87 tests passed
Production build	npm run build passed
Native engine check	npm run engine:check passed
README screenshots	npm run screenshots:readme passed
Packaging	macOS, Windows, and Linux electron-builder targets completed
Native addon resources	Packaged apps contain only the expected platform native addon directory
Packaged app smoke	macOS packaged app launched from release/mac-arm64/ASR Pro.app
Native transcription smoke	Real WAV transcribed through whisper-tiny-en using the native engine
Debian package validation	.deb rebuilt with GNU ar, verified as Debian package format 2.0

Notes:

• macOS artifacts are signed with the local Apple Development identity.
• macOS notarization is not configured for this repo, so the DMG/ZIP are not notarized.
• ASR-Pro-1.0.2-SHA256SUMS.txt contains checksums for all uploaded packages.