Skip to content

Bunyi for Mac 1.3.0

Choose a tag to compare

@github-actions github-actions released this 14 Sep 11:46
· 62 commits to main since this release
5f5cb4e

Bunyi 1.3.0 makes long passages dependable, brings the complete Bunyi CLI to
the Mac, and closes the gaps found in a full keyboard, VoiceOver, and feature
specification audit.

Long passages finish

  • Long text is generated in bounded sections instead of exhausting the model. Bunyi keeps the selected speaker, style, or cloned voice consistent, joins the sections into one recording, and writes one History entry only after the whole passage succeeds.
  • A failed section can recover. Bunyi retries with a smaller section and explains which part it is working on. Stop remains available throughout.
  • Voice Design carries its designed voice into the continuation. It creates one short opening, then uses Voice Clone to preserve that voice through the rest of the passage. The first use may therefore download the separate clone model.
  • Short speech finishes on the model's real end-of-speech signal. Bunyi no longer cuts off valid speech because a text-length estimate guessed too low, and it rejects genuinely unfinished output instead of saving it as though it were complete.

Voice Clone stays local and aligned

  • Automatic transcription now runs locally with Whisper. Auto detects the language without sending the recording to Apple or any other service. The multilingual transcription model is about 148 MB and downloads once on first use.
  • Bunyi samples the same leading, at-most-ten-second window for transcription and cloning. Longer recordings remain valid; the transcript and reference audio can no longer drift apart and leave generation running without useful speech.
  • Old saved voices with a missing transcript repair themselves. Selecting one and pressing Generate recreates and saves the transcript locally. No manual transcript or language selection is required.
  • Missing or mismatched clone input stops before generation. Bunyi identifies the field that needs attention instead of loading a multi-gigabyte model first.

A complete native Mac CLI

  • The release now includes a separate signed and notarized Apple-silicon CLI archive. It exposes generation, transcription, speakers, saved voices, History, playback, models, configurations, Doctor, logs, and backup/restore without opening the desktop app.
  • Persistent local server and detached jobs are included. Automation can keep one MLX model loaded, follow structured progress, cancel work, and recover job results through an owner-only Unix socket. It opens no network listener.
  • The CLI is self-contained. Its archive carries the shared runtime, MLX Metal resources, shell completions, licence, credits, and command reference.

Downloads explain themselves

  • First-use model setup shows measured bytes, speed, time left, verification, and separate file/overall progress. Waiting, stalled, retrying, and resumed transfers are named instead of looking frozen.
  • Interrupted downloads resume safely. Partial files are verified before use, and a paused or unavailable Bunyi model service offers an explicit switch to Hugging Face without discarding completed data.
  • Logs no longer retain credentials embedded in model-source URLs. They keep the host and useful path while redacting secrets.

Keyboard, VoiceOver, and everyday fixes

  • Every window and action has a reliable keyboard route. Focus traversal, reverse traversal, History rows, playback, Settings tabs, Doctor, Logs, dialogs, and the busy-close alert were verified with real key events and VoiceOver.
  • Generation progress is paced for speech. VoiceOver announces meaningful phases and sampled progress instead of being interrupted by every emitted frame.
  • History remains available during generation. It can be inspected while the active generation mode stays safely locked, and Stop remains reachable there.
  • Generate explains missing input in place. The button stays available, then marks and focuses the missing script, voice description, or reference clip without starting model work.
  • Custom models-folder backups now have real progress and cancellation. Stop interrupts the staging copy as well as the archiver and removes only the incomplete new backup.
  • About Bunyi now includes macOS, model credits, software credits, licences, and links. New WAV files also record that they were produced on macOS, while older files remain readable.

Existing downloaded speech models, saved voices, settings, and History remain in
place when updating from Bunyi 1.2.0.


Signed with a Developer ID certificate, notarized by Apple, and stapled.

Requires macOS 15 or later on an Apple silicon Mac.

Open the .dmg and drag Bunyi to Applications. The .zip is the same app for anyone scripting an install.

The first generation in each mode downloads that mode’s model (about 1.5–4.5 GB) with a progress bar; after that it runs offline.

SHA-256:

2f07f699f7c4d66b0ae7a988759c6b1a296bee73b33d9a8e4fd796288efc89a8  Bunyi-1.3.0-5.dmg
2a47c1024980a9599c7a61c6dddc11d47cfe002089ca9993abd039e24c916a2e  Bunyi-1.3.0-5.zip
dcba411d984e2920c54454dc22f9a373bc4faae63adc2afd9e0c1e82ae1e72e0  Bunyi-CLI-1.3.0-5-macos-arm64.zip