Murmur

A Mac app that transcribes audio and video on your Mac and shows who spoke when on a timeline. Nothing you open is uploaded.

Website · Download

• Transcript with speakers. Voz writes the words, Who tells the speakers apart, Align tightens the word timing, Ear checks the spoken language first, and Tongue checks the text afterwards.
• Timeline. A speaker-colored waveform, one lane per speaker, an overview strip and zoom. Click to jump, drag to select.
• Marks. Press M at the start of a section and again at its end, or drag a selection and press M, to keep a named mark. Copy or export one mark or all of them.
• Redaction. One toggle hides names, phone numbers, emails, addresses and IDs (Desert Ant Redact) in the transcript, copies, exports and audio clips (bleeped). Show any single detail again from the sidebar.
• Your colors. Pick the speaker colors in Settings: Vivid, the professional Ink, Carbon, Tableau and Slate, Calm, Color-blind safe (Okabe–Ito), High contrast or Earth, each in light and dark. Every scheme is tested for contrast (≥3:1) and for speaker colors that stay distinct (OKLab distance).
• Editing. Fix words in place (edited words are re-timed against the audio), rename, merge and reassign speakers. Everything is undoable.
• Export. Plain text, text with timestamps, Markdown, SRT, WebVTT, word-level JSON, and audio or video clips (one file per mark, or joined).
• Meeting recording. Records your microphone ("Me") and the sound of other apps ("Others") as two tracks, with a live transcript in a small floating window that is hidden from screen sharing. ⌥⌘M marks a moment from any app.
• Extras. Uhm finds um/uh (hidden in exports by default), Clear enhances the audio for playback, transcription and export, and Apple Intelligence names the project and writes a summary (shown in the sidebar and included in Markdown and JSON exports).
• Languages. Ear names the spoken language; when it is unsure, Murmur reads a sample with Voz and lets Tongue decide. Languages Voz lacks can go to Apple's on-device transcriber instead.

Requires macOS 26 or later on Apple silicon (arm64 only: the Desert Ant SDK does not build for Intel).

In the browser

Murmur in the browser does the same for a file you open, in Chrome, Edge or Safari 26+, without uploading it. Voz, Who, Ear, Redact and Clear run in the page (WebGPU and WebAssembly); their weights download once from Hugging Face and stay in the browser. Projects stay in the browser too. Compared to the Mac app it is slower (an hour takes 12–15 minutes) and records only the microphone and browser tabs, not other apps, and leaves out Align (word re-timing), Uhm (fillers), summaries and video clips. The code is in web/app; the plan and the measurements behind it are in docs/web-plan.md and web/spike.

cd web/app && npm install && npm run dev    # http://localhost:5200
npm test                                     # the TypeScript core against MurmurCore's fixtures

On iPhone and iPad

The MurmurMobile target (iOS 26) shares the pipeline, models and project format with the Mac app: open a file or record with the microphone (live transcript, moment marks, keeps recording with the screen locked), then the timeline, transcript, marks, editing, redaction and sharing work as on the Mac. Recording other apps' sound is Mac-only. Not released yet; the plan and what is left are in docs/ios-plan.md.

make generate && open Murmur.xcodeproj    # scheme MurmurMobile

Building

brew install xcodegen
make run        # Debug build and launch
make test       # MurmurCore unit tests
make build      # Release build -> build/Murmur.app
make install    # Release build into /Applications
make dmg        # build/Murmur-<version>.dmg

On first launch Murmur downloads its speech models (about 515 MB) into ~/Library/Application Support/com.djui.Murmur/Models. Optional models (Uhm, Clear) download when turned on.

Headless

The app binary can run the pipeline without UI, which is handy for testing:

build/DerivedData/Build/Products/Debug/Murmur.app/Contents/MacOS/Murmur -process talk.m4a -title YES -out talk.json

Flags (AppKit-style -key value pairs): -language xx, -enhance YES, -fillers NO, -title YES, -stems mic.m4a,system.m4a, -out file.json. The transcript goes to stdout (logs to stderr); -format plain|timestamps|markdown|srt|vtt|json picks its format (default timestamps).

Other checks: -live mic.m4a,system.m4a streams two tracks through the live transcript, -clip media -ranges 7-14,20-28 -out folder [-joined YES] [-video YES] cuts clips, and -mix mic.m4a,system.m4a -out recording.m4a mixes stems the way a finished recording does.

On an M2 (Release build), ten minutes of audio take about 10 s: decode 0.6 s, language 1.5 s, Voz 1.7 s, Who 1.6 s, Align 1.5 s, Uhm 3 s. Debug builds are much slower because the SDK's DSP code is unoptimized; measure with make build.

Releasing

make release bumps the version, builds a DMG, generates the Sparkle appcast and publishes a GitHub Release. The Sparkle signing key lives in the release machine's login keychain (its public half is SUPublicEDKey in Murmur/Info.plist); back it up with generate_keys -x, since updates can't be signed without it.

Releases are notarized once these exist; until then they are signed with an Apple Development identity and only run on the developer's Mac:

1. A Developer ID Application certificate in the keychain.
2. xcrun notarytool store-credentials murmur and export NOTARY_PROFILE=murmur.

Layout

• Packages/MurmurCore — the project model, merging words with speakers, paragraphs, text formats, editing and waveform peaks. Pure Swift, unit tested.
• Murmur/ML — model downloads and the processing pipeline (Desert Ant SDK, Apple's transcriber, Foundation Models).
• Murmur/Document — the .murmur package (project.json, peaks.bin) and the per-window session.
• Murmur/UI — library, project window, timeline, transcript, sidebar, export, onboarding, settings.
• Murmur/Recording — microphone and system-audio capture, live transcript, the compact panel.

Credits and licenses

Speech recognition by Voz, based on NVIDIA Parakeet TDT 0.6B v3 (CC BY 4.0). Who, Ear, Align, Uhm, Clear and Tongue by Desert Ant Labs under the Desert Ant Labs Source-Available License (free up to 100k monthly active devices). Loading a Desert Ant model reports an anonymous usage count to Desert Ant Labs for licensing. Updates by Sparkle.