Skip to content

OpenWhisperer 1.5.0 — Fully Native

Choose a tag to compare

View all tags
@PerIPan PerIPan released this 24 Jun 21:48
v1.5.0
8e147e5

What's New in 1.5.0

  • Fully native, no Python — speech-to-text (WhisperKit) and text-to-speech (FluidAudio Kokoro) run in-process on the Apple Neural Engine. The Python server, virtualenv, and setup.sh are gone — install is just "drag to Applications," cold start is faster, and a whole class of dependency-drift failures disappears.
  • In-app streaming playback + instant barge-in — replies start speaking after the first sentence and play gaplessly; saying "hold on" (or starting a new turn) stops audio and cancels in-flight synthesis in-process.
  • Tagless voice mode — no more [VOICE:] tag in CLAUDE.md. The app fingerprints each dictation and a UserPromptSubmit hook routes the spoken reply to the session you dictated into; only dictated turns are spoken.
  • Warm redesign — the menubar and transcription overlay now match openwhisperer.com: a warm cream/gold palette with a Fraunces serif wordmark, in full light and dark mode.
  • WhisperKit 1.0 — speech-to-text engine updated to the 1.0 stable release.
  • Reliability hardening — generation-guarded TTS cancellation, loopback TTS server body-size cap + surfaced bind-failure, the "Speaking…" lock clears if the output device drops mid-reply, and a uniform voice-turn freshness window.
  • Garbled-speech fix on Apple Silicon — on some chips (notably M3 / macOS 15) FluidAudio 0.15.4 mis-read a strided CoreML array and produced fluent-but-wrong, "foreign-sounding" speech regardless of the text. Updated to the upstream fix so synthesis is correct across Apple Silicon generations.
  • Delete downloaded models — a maintenance button in Server & Logs clears the STT/TTS model caches after a confirmation that shows exactly how much space it frees; the models re-download automatically on next use. It's also the quickest recovery if a model cache ever gets wedged.

First launch downloads the models (~1.7 GB) and compiles them for the Neural Engine (~1–2 min, one-time). Every launch after is instant.

Install

Download OpenWhisperer-1.5.0.dmg below and drag it to Applications. The app isn't signed with a paid Apple Developer certificate, so allow it to run:

xattr -cr /Applications/Open\ Whisperer.app

On first launch macOS asks for Microphone, Accessibility, and Speech Recognition permissions.

Credits

The native rewrite — in-process WhisperKit STT + FluidAudio Kokoro TTS, streaming playback + barge-in, and the tagless voice-turn handshake — was contributed by Hakan Ensari. Thank you!

Assets 3
Loading