Image: GitHub
Why it mattersA developer who speaks a long comment or a chat reply through a cloud dictation service is sending that audio to a third party, and now has an on-device option to compare it against.
A developer who dictates code comments, chat replies or an email to a colleague through a cloud service is sending that audio to a third party. BetterWispr, an open-source Mac dictation app that reached 402 stars in the 12 days since its first commit, offers an on-device option: hold Option-Space, speak, release, and the text lands in whatever app you were typing in.
The project, by opennookorg, is published under Apache-2.0 and names itself in its README as "an open-source alternative to Wispr Flow". It was created on 2026-10-07 and pushed most recently on 2026-10-15. The code is Swift and the download is a Mac disk image.
What runs where
Four speech-to-text engines are listed in the README. Three run on-device: Apple's own speech framework (needs no download), NVIDIA Parakeet v3 (covering 25 European languages), Parakeet v2 for English, and Whisper large-v3 turbo. The project says audio never leaves the Mac unless the user adds a cloud connection.
The app can also route to a cloud provider, listed in the README as Sarvam AI (saaras:v4), Smallest AI (pulse), or any endpoint that speaks the OpenAI transcription API. Keys live in the macOS Keychain. The README is clear that when a cloud connection is in use, that provider's own pricing and data policies apply, and the app does not fall back to the cloud on its own.
Vocabulary, meetings and connected apps
A few features sit on top of the raw transcription. The vocabulary list learns a spelling hint once a user corrects a misheard word in the history pane. Voice commands include "comma", "new line", "scratch that", "sorry, remove that", and "at the rate KV" for @KV. A meetings mode produces a live transcript of both sides of a call plus a summary and action items, running through Apple Intelligence or a local Ollama model.
The feature a working developer may notice first is Connected apps. The README says it exposes a read-only MCP server named BetterWispr --mcp on the local Mac, which lets other apps that speak MCP list, search and read the user's meeting notes and transcripts. It is off by default. When it is on, the app the user connects reads through that MCP server and sends whatever it reads to its own AI provider, so turning it on is still an act of sharing data.
What it does not do
Three things the README states as limits. BetterWispr requires macOS 14 or later and runs best on Apple silicon. The workspace file at ~/Library/Application Support/BetterWispr/workspace.json is not encrypted, so secrets would not belong in it. And if a paste cannot safely reach the previous app, the dictation is copied to the clipboard instead, which is a different user experience that the owner should know about.
For a team paying for Wispr Flow seats, the practical question is whether on-device Parakeet or Whisper transcribes the team's own vocabulary well enough. The project ships a model picker, so the test is one install.
Source
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.
