BetterWispr turns speech into text in the Mac app you are already using. Its first public release shipped October 8; downloadable releases include Apple Silicon and Intel builds. It is a concrete option for dictating coding-agent prompts, email or notes without a mandatory transcription subscription—not a voice-controlled agent that operates your computer.
- Setup and interface: Install the DMG, grant Microphone and Accessibility permissions, and select Apple on-device speech or download a Parakeet or Whisper model. Hold Option–Space, speak, then release to transcribe and paste. Blocked insertion falls back to copying the text; transcription is not streamed word by word while you dictate.
- Technical vocabulary: Add spelling hints and phrase replacements for names and jargon. Cleanup can keep recognized words unchanged, remove English fillers and repetitions, or use a selected notes model to edit for clarity. Raw transcripts remain recoverable from History.
- Meeting notes: Capture microphone and other apps’ audio, then generate summaries and action items with local Ollama or Apple Intelligence. System-audio capture needs macOS 14.2+, and Apple Intelligence summaries need supported hardware running macOS 26+. A transcript can be saved without a summarization model.
- Cost and access: The app is free, Apache-2.0 licensed, and needs no account or API key for built-in local speech. macOS 14+ is required; Apple Silicon is recommended. Optional external providers have their own pricing and model weights retain their own terms.
The important limitation is that local-first is a configuration, not a blanket privacy guarantee. Choosing a remote speech connection sends audio to that endpoint; choosing Claude Code, Codex or a remote notes API can send transcripts out for summaries or stronger cleanup. History is enabled by default and is not encrypted by the app. Turn it off for sensitive dictation and choose local speech and notes providers deliberately.
This is a source-based announcement capture, not a hands-on accuracy or latency test. The initial release also says physical Intel hardware and minimum-supported-macOS testing were still pending. Start with a short sample of your own technical vocabulary before trusting the output in an agent prompt.