Why it is worth attention

It offers unlimited, private voice dictation without cloud or subscription, using Apple's on-device models and optional Whisper engine, with pronunciation learning and text cleanup.

Who it is for

Privacy-conscious professionals
Writers and journalists needing offline dictation
Developers who prefer open-source tools
macOS power users seeking customizable dictation

Use cases

Hands-free text input in any macOS application
Dictation with custom vocabulary (e.g., medical or technical terms)
Quick text formatting, cleanup, and tone rewriting
Transforming rough spoken notes into structured prompts

Strengths

100% on-device processing ensures complete privacy
Two offline engines: instant Apple or precision Whisper
Pronunciation learning that adapts to your speech over time
Integrated cleanup pipeline, per-app styles, and text transforms

Considerations

Requires macOS 26 (Tahoe) and Apple Silicon Mac
Styles and Transforms depend on Apple Intelligence being enabled
Initial setup involves building from source and granting permissions

README quick start

Murmur 🎙️

Private, unlimited voice dictation for macOS — 100% on-device.

Hold fn, speak, release — clean text appears at your cursor in any app. No cloud, no subscription, no word limits. Your audio and transcripts never leave your Mac.

Murmur is an open-source, fully local take on the modern AI dictation app (in the spirit of Wispr Flow), built natively in Swift on Apple's on-device speech and language models, with an optional local Whisper engine.

Features

Push-to-talk dictation — hold fn (or right ⌥) anywhere; release to paste at your cursor. Double-tap for hands-free mode.
Two recognition engines, both offline:
- Apple — instant, built into macOS (SpeechAnalyzer, macOS 26).
- Whisper — optional precision engine via WhisperKit (CoreML on the Neural Engine). Your vocabulary is fed into the decoder prompt.
Pronunciation learning — a Voice Training page learns how you say tricky words; corrections you make to transcripts are diffed and learned automatically; everything biases future recognition.
Cleanup pipeline — filler-word removal, spoken "new line"/"new paragraph", auto-capitalization, personal dictionary, snippets (say a trigger phrase → paste a saved block).
Styles — per-app tone rewriting (formal / casual / very casual) using Apple Intelligence's on-device model.
Transforms — select text in any app, press ⌥1 to polish grammar or ⌥2 to turn rough notes into a structured AI prompt, rewritten in place.
Dashboard — history with search and correction-learning, usage stats (words, WPM, day streak), insights chart, a Voice Profile persona derived locally from what you dictate, scratchpad.

Requirements

macOS 26 (Tahoe) or newer
Apple Silicon Mac
Xcode 26 command-line tools (xcode-select --install)
For Styles / Transforms / Voice Profile: Apple Intelligence enabled
For the Whisper engine: a one-time model download (150 MB – 1.6 GB)

Build & run

git clone 
cd murmur
./scripts/make_app.sh     # builds build/Murmur.app
open build/Murmur.app

Optional: run ./scripts/make_signing_cert.sh once to create a local self-signed signing certificate — this keeps macOS permission grants valid across rebuilds. Without it the app is ad-hoc signed and you'll need to re-grant Accessibility after each

Description

Private, on-device voice dictation for macOS

Related repositories

Similar projects matched by category, topics, and programming language.

FrigadeHQ

yap

Yap is a blazing-fast, on-device macOS voice dictation app that transcribes speech into any text field via a global shortcut, using Apple's macOS 26 Speech APIs with no network calls or model downloads.

Swift

197

antiwork

chromeless

Chromeless is a minimal macOS browser built on WKWebView that removes all interface chrome, providing a distraction-free window and a CLI screenshot tool.

Swift

267

drumih

turbo-fieldfare

TurboFieldfare is a custom Swift and Metal runtime that runs the Gemma 4 26B-A4B instruction-tuned model in about 2 GB of RAM on Apple Silicon Macs, including 8 GB models, by streaming experts from SSD.

Swift

256