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.

• Push-to-talk dictation — hold fn (or right ⌥) anywhere; release to paste at your cursor. Double-tap for hands-free mode.
• Four recognition engines, all offline:
  • Apple — instant, built into macOS (SpeechAnalyzer, macOS 26).
  • WhisperKit — precision engine via WhisperKit (CoreML on the Neural Engine). Your vocabulary is fed into the decoder prompt.
  • whisper.cpp — the same Whisper models, running via Metal on the GPU instead, with phrase-based filtering for the sign-off hallucinations ("Thank you.", etc.) Whisper-family models are known to produce on near-silent audio.
  • Parakeet — NVIDIA's model via FluidAudio (CoreML on the Neural Engine). Murmur's fastest engine; no vocabulary biasing yet.
• Harper grammar pass — Harper runs alongside the Apple Intelligence edit pass below: deterministic, millisecond-speed local linting (agreement, punctuation, repeated words) that catches what an LLM edit occasionally misses. No model, no warm-up, no network.
• Per-app profiles — one place to set what Murmur does when you dictate into a specific app: tone and note template together, instead of two separate override systems answering "what happens in Slack?" differently.
• Ask Murmur — ask questions about your own dictation history, answered entirely on-device via lightweight keyword retrieval into the local model (no embeddings, no network).
• Note templates — restructure a transcript into a specific document shape via the on-device LLM; trigger one by voice at the start of a dictation, or apply manually.
• 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.
• Guided first run — welcome → permissions → recognition engine → hotkey → mic test, shown once.

• 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)

git clone <this-repo>
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 rebuild.

1. Microphone — allow when prompted on first dictation.
2. Accessibility — allow when prompted (needed for the global hotkey and for pasting). If the app still shows it as missing, use Settings → Reset Grant & Relaunch inside Murmur.

.build/debug/Murmur --selftest                          # formatter + learning tests
.build/debug/Murmur --transcribe audio.wav              # Apple engine
.build/debug/Murmur --transcribe audio.wav --engine whisper
.build/debug/Murmur --format "um hello new line hi"     # cleanup pipeline only
.build/debug/Murmur --transform "fix this grammer pls"  # on-device LLM polish

Everything runs on this Mac: recognition (Apple SpeechAnalyzer or a local Whisper/Parakeet model), cleanup, tone rewriting (Apple Intelligence), and the Voice Profile analysis. Murmur makes no network requests except the one-time model downloads by macOS itself (Apple speech assets) and, if you opt into a non-Apple engine, that model's own one-time fetch (Hugging Face for Whisper, FluidAudio's own CDN for Parakeet). Dictation data is stored only in ~/Library/Application Support/Murmur/.

Swift Package. WhisperKit and FluidAudio are remote SwiftPM dependencies, each used only if you pick that engine; whisper.cpp/ggml and Harper are vendored as prebuilt xcframeworks under Vendor/ (Harper's Rust source is included, whisper.cpp's isn't — see NOTICE.md for both).

HotkeyMonitor  →  AudioRecorder  →  Transcriber (Apple / WhisperKit / whisper.cpp / Parakeet)
                                        ↓
     TextFormatter → LearnedStore → SnippetStore → RewriteEngine (Styles) → HarperChecker
                                        ↓
                        TextInserter (clipboard + ⌘V)

See PLAN.md for the original design document and CHANGELOG.md for what's new in each release.

MIT. Not affiliated with Wispr Flow, OpenAI, NVIDIA, or Apple.