Glonce

BetterWispr offers free open-source Mac dictation that runs locally

toolsThe Glonce editorsHacker News🔥54Glonce rating54Impact35Novelty30Relevance80Surprise35Players10Credibility70

BetterWispr is a dictation app for Mac, shown on its own site as a Show HN launch. It is free and open source under Apache 2.0, requires macOS 14 or later, and recommends Apple Silicon for local model inference. The site says it needs no dictation subscription, no account and no API key.

The core idea is that speech stays on the machine. Built-in models transcribe on the Mac, there is no cloud transcription fallback, and models download once and then work offline. There are nine built-in models drawn from Apple speech, Parakeet and Whisper. Parakeet covers English, Japanese and 25 European languages (Parakeet TDT v3 handles the European set, a separate Parakeet model handles Japanese), and Whisper covers more. Users who prefer a cloud service can instead choose their own Sarvam AI, Smallest AI or OpenAI-compatible API.

Usage is push-to-talk. The default is to hold Option + Space in any app, speak, and let go; the user can switch to press-to-toggle or record another shortcut using Command, Option, Control or an F key. Transcription begins after recording stops, so text does not appear word by word while the user talks. The result is inserted into the app the user was in. If pasting is blocked, the text is copied to the clipboard so nothing is lost; secure fields and some apps may block automatic insertion.

Cleanup is configurable. Light, the default, drops English ums, repeats and false starts. Medium also edits for clarity using the chosen notes model, running on the Mac through Apple Intelligence or Ollama. None keeps every word. Per-app styles adjust punctuation and capitalisation: casual drops the final period, very casual drops capitals too, and formal leaves the words alone. Filler cleanup, styles and voice commands (say comma, question mark or new paragraph; say scratch that to delete the last sentence) work in English only.

A Notetaker mode transcribes the microphone as Me and, on macOS 14.2 or later, other apps' audio as Them, next to the user's own notes. Afterwards Apple Intelligence on supported Macs running macOS 26, or a local Ollama model, writes the summary, decisions and action items on the Mac. Claude Code, Codex and OpenAI-compatible endpoints are optional and send the transcript to that provider. Without a notes model, the transcript and notes are still saved.

The app also keeps a local history with raw and corrected text, and Use Original restores exactly what was said; history can be turned off. Fixing a word in History or right after a paste adds it to Vocabulary so the app spells it that way next time, and learning can be turned off in Settings. A dashboard shows words spoken, pace, words saved by cleanup, which apps are dictated into and a daily streak, all counted from history on the Mac. Updates arrive from GitHub Releases through Sparkle and install only when their EdDSA signature matches. The source can be read on GitHub and built with make dev.

Key facts

  • BetterWispr is a free, Apache 2.0 licensed dictation app for macOS 14 or later; Apple Silicon is recommended for local inference.
  • Nine built-in models from Apple speech, Parakeet and Whisper transcribe on the Mac, with no account, no cloud fallback and offline use after a one-time download.
  • Parakeet covers English, Japanese and 25 European languages; Whisper covers more. Filler cleanup, styles and voice commands work in English only.
  • Transcription starts after recording stops rather than appearing live, and the default shortcut is holding Option + Space.
  • Notetaker labels the microphone as Me and other apps as Them, then writes a summary, decisions and action items using Apple Intelligence on macOS 26 or Ollama.

Why it matters

Dictation tools often send audio to a server. BetterWispr's pitch is the opposite: built-in models run on the Mac, there is no cloud transcription fallback, and the app needs no account or API key. It is also open source under Apache 2.0, so the code can be read on GitHub or built with make dev. The site frames this as local by design: downloads happen when the user chooses to install models, and speech does not need a round trip to a transcription server.

Who it affects

Mac users on macOS 14 or later who want dictation in any app without a subscription, and who care about keeping speech on their own machine. Apple Silicon is recommended. English speakers get the most, since cleanup, tones and voice commands are English only. Users of other languages can rely on Parakeet for Japanese and 25 European languages, or on Whisper, and the site advises testing the chosen model with your own speech. People who take meeting notes are the audience for Notetaker, which needs macOS 14.2 or later to capture other apps' audio.

How to use it

Install the app and follow the welcome guide through permissions and the first local speech model. Microphone permission lets it record; Accessibility permission allows automatic insertion into another app; Apple speech also needs Speech Recognition permission. Then place the cursor in a text field, hold Option + Space, speak, and release to transcribe. Clean text lands in the app you were using. Shortcuts, press-to-toggle mode, cleanup level (Light, Medium or None) and per-app styles are set in the app. The source can be built with make dev. Cloud options such as Sarvam AI, Smallest AI or an OpenAI-compatible API are available only if the user picks them.

How solid is it

Everything here comes from the product's own site, so these are the project's claims rather than independent findings. What is concrete: the Apache 2.0 licence, the nine built-in models from Apple speech, Parakeet and Whisper, the macOS version requirements, and signed updates that install only when the EdDSA signature matches. The source says no author, developer or company behind BetterWispr is named, and no release date, version number or changelog is given.

Risks and caveats

No accuracy, latency or speed benchmarks are given for any model, and no comparison results with other dictation apps are stated on the page. Transcription happens after recording stops, so there is no live text. Cleanup, tones and voice commands work in English only. Language support depends on the selected model, Apple speech depends on locale and installed assets, and model availability and performance depend on hardware and language. Automatic insertion can be blocked by secure fields and some apps, in which case the text goes to the clipboard. Medium cleanup and Notetaker summaries stay on the Mac only with Apple Intelligence or Ollama; Claude Code, Codex and API connections send the text to that provider. Model files use storage on the Mac, and the source gives no size figures.

“Built in models transcribe your speech on your Mac, with no account and no cloud fallback.”

— BetterWispr website