WhisperDrop 2.5.0

WhisperDrop 2.5.0 has one place to choose the model, live text that keeps everything you said, and dictation that can write into the text field while you speak. It also includes the experimental Parakeet v3 engine and saved-note renaming, prepared for 2.4.0 but not published separately. Whisper Turbo Q5 remains the default.

Changes since 2.3.0

One place for models

• Settings > Models is the only place to download, delete and choose a model. Before, the main window chose the model for files and Settings chose it for dictation and notes, and the two could differ without warning. The main window and the sidebar now show the model and open Settings > Models.
• The model you had chosen in Settings becomes the model for everything.
• Advanced: a different model for each task can give dictation, notes or files their own model, for example Parakeet v3 for dictation and Whisper for lectures.
• Keep model ready in the menu bar and the Unload model setting were two controls for the same thing. Keep model ready is now one of the Unload model choices; turning it off in the menu bar restores your previous interval. Keep audio has one name beside New note and in Settings.
• Notes and dictation check that their model is downloaded before the microphone opens. Before, a note could record with a missing model and skip every part of the transcript.

Parakeet v3, experimental

• Download Parakeet TDT 0.6B v3 with a 4-bit encoder, about 489 MB. It runs locally in a native Swift helper using MLX and supports 25 European languages, detected automatically. It ignores language hints and vocabulary lists. Unsupported language selections are refused before capture starts.
• Parakeet file imports use pause-aligned windows of up to 35 seconds, in a process separate from the dictation and note model.

Live text, with Parakeet

• Notes show every word heard since the last confirmed paragraph, in grey. Before, the grey text covered only the last 20 seconds, so earlier unconfirmed words disappeared during long passages. The transcript follows new text only while you are at the bottom; scroll up to read earlier text and use Latest to return.
• Dictation can write into the text field while you speak (Settings > Dictation > Live text > In the text field). Native text fields show every word and correct it in place; when you finish, the whole dictation is decoded once more and the field is corrected to that text. Browsers and apps whose fields ignore accessibility edits receive words once they settle, about six seconds behind, and the newest words show in the dictation bar; there the typed words stay and only the rest is decoded at the end. Press Escape to discard the dictation, including the words already written. If you click elsewhere or type while dictating, WhisperDrop stops writing and copies the full text to the clipboard instead.
• The dictation bar shows up to four lines of live text instead of one truncated line.
• Live requests stay short however long you speak: words settle once about six seconds of audio follow them, and later requests start ten seconds before the settled words, so each request covers at most 30 seconds.

Other changes

• Dictation that starts right after other text, such as "end." with the caret after it, begins with a space instead of joining the previous word.
• Rename a saved note with Rename… beside its title or from its library context menu.
• Whisper dictation keeps the full encoder window. The shortened window of 2.3.0 returned unrelated words and duplicated sentences on short clips.

Validation and limits

• 114 standard tests pass in debug and release builds; seven model replays that need local recordings are skipped by the standard suite and were run separately.
• Live text was replayed silently through the real model host on Italian and English recordings cut into dictations of about one minute. Against decoding each dictation whole, word error rate went from 7.22% to 7.36% on Italian audiobooks, 4.91% to 5.11% on Italian TED talks and 1.41% to 1.28% on English audiobooks. The wait after a dictation ended dropped from about 1.7 to 2.2 seconds to about 0.6 seconds when only the unsettled end was decoded. Details are in docs/parakeet-evaluation.md.
• A 4-minute note replayed at recording pace kept every sample; its grey text grew to 121 words, 56 seconds of speech not yet confirmed.
• Dictation into the text field was checked with recorded clips in TextEdit, the Safari address bar, Safari and Chrome web text areas, after existing text, with the caret moved mid-dictation and with Escape. It was not tested with a live microphone, a physically held key, Electron apps (Slack, Notion, VS Code), Microsoft Word or terminals. Password fields turn live writing off.
• In browsers, words appear about six seconds after you say them, and a slow request can make them later.
• The packaged app transcribed an Italian clip with Parakeet v3 and with Whisper Turbo Q5.
• Tested only on macOS 26 on Apple Silicon.

Install

Apple Silicon, macOS 14 or later. Quit the previous version and replace the app from the DMG. Models, notes and settings are stored separately from the application bundle.

This build has an ad hoc signature and is not notarized. Gatekeeper may block a downloaded copy. If you choose to trust the build, use Apple's Open Anyway flow in System Settings, Privacy & Security. Because the signature changes with every ad hoc build, macOS may ask again for Accessibility, Input Monitoring and Microphone access after updating; if dictation does nothing, remove WhisperDrop 2 from those lists in Privacy & Security and allow it again.

The app grows from about 144 MB in 2.3.0 to about 477 MB, because it includes Parakeet and two MLX shader libraries. It requires no Python or cloud service at runtime.

Release assets include SHA256SUMS.txt and the unchanged FFmpeg 8.1.2 source archive. Runtime license notices are bundled in the app.