What is it?
BetterWispr is a dictation app for macOS. Hold ⌥ Space, talk, let go, and the text is pasted into whichever app had the cursor. Speech is transcribed on the Mac with Whisper, NVIDIA's Parakeet or Apple's built-in recognition, and a notetaker covers meetings as well.
The comparison of local speech-to-text tools places it among dictation apps, meeting notetakers and transcription servers, and tracks what each one sends off the machine.
Why is it interesting?
- No silent cloud fallback: built-in models only reach the network to download, and only when you ask. An API connection is used when you select it, never as a backup for a local model.
- Several model families: Apple's speech recognition with nothing to download, Parakeet TDT v3 for 25 European languages, Parakeet v2 for English, Whisper large-v3 turbo through WhisperKit.
- Vocabulary that learns: spelling hints and replacements for names and jargon. Fix a misheard word once, in History or right after the paste, and it comes out right the next time.
- Meeting notes: your microphone is transcribed as "Me" and, from macOS 14.2, other apps' audio as "Them". Apple Intelligence or a local Ollama model writes the title, summary, decisions and action items.
- Published measurements: the changelog puts numbers on model choices. On 300 English video-call clips, Parakeet Ultra, a further-trained TDT v3 recommended since 0.1.5, got 14.45% of words wrong against 15.96% for Parakeet TDT v3. Accuracy on real dictation is not measured, since the app keeps no dictation audio.
Good to know
Mac only, macOS 14 or later. The Apple Silicon and Intel DMGs weigh about 15 MB, models not included; Apple Silicon is recommended. Some features need a newer system: macOS 14.2 to capture other apps' audio in meetings, macOS 15 for speaker separation in 0.1.5, macOS 26 for Apple Intelligence summaries. Off the Mac, Murmure does dictation on Windows and Linux with the same Parakeet TDT v3, and Meetily transcribes meetings on Windows.
The extras are English-only. Transcription covers whatever languages the chosen model supports, but voice commands ("comma", "new line", "scratch that"), filler cleanup and per-app writing styles only work in English, according to the site's FAQ. The same goes for rewriting with a language model (the Medium cleanup level) and for the Parakeet vocabulary booster.
Local by default, with exits you opt into. An API connection (Sarvam AI, Smallest AI or OpenAI-compatible) receives your audio, and its keys are kept in the macOS Keychain. Handing summaries or Medium cleanup to Claude Code, Codex or an API sends the text to that provider. Settings, vocabulary and history are stored in an unencrypted workspace.json file, and the app checks GitHub for updates.
Positioned as a Wispr Flow alternative. The README calls BetterWispr an open-source alternative to Wispr Flow. The app is free, Apache-2.0 licensed, needs no account or subscription, and takes donations through Buy Me a Coffee.
A repository opened on 7 October 2026. Six releases followed between 8 and 12 October, up to 0.1.5, still the latest published on 26 October. The README describes the main branch: transcribing files and links, and sharing meetings over MCP with Claude or Cursor, belong to a 0.1.6 that the changelog dates 15 October and that no release ships yet. One developer authored 273 of the 279 commits.