Murmur

Murmur is a free desktop speech-to-text app for macOS and Windows (beta). Press a shortcut, speak, and Murmur pastes the transcription into the active application.

Murmur transcribes on your computer: the audio stays there, there is no quota, and no account is needed. Sign in with a ChatGPT account from Murmur to add AI cleanup, dictation styles, more languages and cloud transcription. It requires no API key and no separate Codex app.

Requirements

macOS

• macOS 10.15 or later
• A working microphone
• A ChatGPT account, signed in from Murmur, for cloud transcription, AI cleanup and meetings. Local transcription, the default, runs without it. A Mac that cannot run the local model needs it to finish setup.
• Accessibility permission for global shortcuts and text insertion
• For local transcription: macOS 15 or later on Apple silicon, or an Intel Mac, and a one-time download of about 500 MB (740 MB on Intel)
• For meetings with system audio: macOS 14.4 or later on Apple silicon, and the System Audio Recording permission
• Network access for cloud transcription and AI cleanup

Windows (beta)

• Windows 10 or 11, x64
• A working microphone
• For local transcription: a one-time download of about 740 MB
• A ChatGPT account, signed in from Murmur, for cloud transcription and AI cleanup. Local transcription runs without it.
• Network access for cloud transcription and AI cleanup

Murmur keeps the ChatGPT session in the file the Codex apps use, so both share one sign-in: CODEX_HOME/auth.json, or else ~/.codex/auth.json on macOS and %USERPROFILE%\.codex\auth.json on Windows. When the access token is about to expire, or when the service rejects it, Murmur refreshes the session the same way the Codex CLI does and writes the new tokens back to that file. It changes only the token fields and last_refresh, keeps the file private (mode 600), and replaces it in one step. If Codex updated the file while Murmur was refreshing, Murmur keeps the Codex version. When the refresh token itself has expired, sign in again from Murmur.

Private interfaces

Murmur uses an internal service rather than a public, documented API: the ChatGPT transcription endpoint that the Codex apps use. It promises no stability, so a Codex or ChatGPT update can break transcription without notice until Murmur is updated. Murmur identifies itself to the transcription endpoint the way the Codex desktop app does. Use of these services stays subject to the provider's terms for your account.

Install

macOS

1. Download the DMG from the latest GitHub release: aarch64 for Apple Silicon, x64 for Intel.
2. Open the DMG and drag Murmur to Applications.
3. Open Murmur. Releases are signed with a stable self-signed certificate but are not notarized by Apple, so macOS blocks the first launch. Open System Settings → Privacy & Security and click Open Anyway. Since macOS 15, right-click → Open no longer bypasses this check.
4. Grant Microphone and Accessibility access when Murmur asks.

Murmur's built-in updater installs releases signed with the same certificate, so macOS keeps both permissions and asks nothing again after an update. Murmur does not store anything in the keychain.

Windows (beta)

1. Open the Windows releases and download Murmur_…_x64-setup.exe from the first one in the list.
2. Run the installer. It is not signed, so SmartScreen blocks it: click More info, then Run anyway. Murmur installs for the current user.
3. Open Murmur, download the speech model or sign in with ChatGPT. Hold Ctrl+Win to dictate, or double-tap it to record hands-free.

The Windows source lives on the beta branch for now, not on main.

Features

• One recording shortcut (Option+Space by default, Ctrl+Win on Windows): hold it to talk, or double-tap it to record hands-free and press it once to stop
• A cancel shortcut (Escape by default), active only while Murmur records or transcribes
• A shortcut (Control+Command+V by default) and tray item that paste the latest transcription again
• Transcription on this Mac or in the cloud, chosen in Settings → Transcription. On the Mac, NVIDIA Parakeet TDT 0.6B v3 covers 25 European languages with no quota. When it can't serve a dictation (model not downloaded yet, failure, another language), Murmur uses the cloud if ChatGPT is connected; when ChatGPT refuses in cloud mode, a downloaded model takes over. History marks each such dictation.
• When no text field is focused, nothing is pasted or copied; the overlay offers a Copy button that puts the text on the clipboard
• Automatic language detection, or an explicit dictation language
• Microphone and output-device selection
• AI cleanup through your ChatGPT account, on by default: removes fillers, applies self-corrections, fixes punctuation, and formats lists and spoken line breaks. Any failure or timeout pastes the raw transcript.
• Dictation styles for cleanup: Formal, Normal (the default) or Casual
• A dictionary of custom words, spelled your way in each dictation
• Names from the frontmost window and a short excerpt around the cursor as hints for AI cleanup, read through Accessibility (UI Automation on Windows) and never stored; password fields and password managers are never read
• Automatic level boost for quiet speech, and a Whisper mode for very quiet speech
• Transcription history with saved recordings, keeping the raw and cleaned text; a failed or unsatisfying transcription can be sent again from its recording
• Insights: total words, words per minute over the last 30 days, current and longest streak with a 20-week heatmap, and usage by app category
• Meetings: record a call (Control+Option+Command+M or the menu bar), then get a transcript with speakers told apart and a summary with a title. See Meetings
• A minimal waveform overlay at the top or bottom of the screen, which can be turned off
• Audio feedback, tray controls, and automatic updates

The Windows beta has dictation, transcription on the PC or in the cloud, AI cleanup with styles, names read from the active window, history, Insights, custom words, shortcut keys shown for your keyboard layout, sounds, and launch at login. It does not have Meetings or automatic updates yet.

Getting started

Setup first offers to transcribe on this Mac. Download and continue starts the model's download in the background, so it runs while you grant microphone and Accessibility permissions; a small pill shows its progress until the model is ready. Use the cloud instead keeps the model off this Mac. A Mac that cannot run the model skips this step.

Then connect ChatGPT: Sign in finishes in your browser. If the Codex app is already signed in on this computer, Murmur uses that session. With local transcription this step is optional: Skip it to dictate without an account, without AI cleanup and the cloud's extra languages. With cloud transcription, ChatGPT must be connected before setup can finish. Until the model is ready, a connected account transcribes in the cloud; without one, dictation waits for the model.

"Configured" means Murmur found a local session that is valid or that it can refresh. A session that expired and has no usable refresh token shows as "Session expired". Configured does not prove that the cloud service will accept the next request, because network or account limits can still prevent transcription.

When a dictation fails, the recording overlay shows the reason for a few seconds. If the overlay is turned off, Murmur posts a macOS notification instead. Murmur retries a dictation twice after a timeout, a dropped connection, or a server error. It does not retry an expired session, a rate limit, or another rejected request. When Codex reports that its dictation usage limit is reached, Murmur resends the recording at once against ChatGPT's own dictation limit, and uses that limit for the next hour. A finished transcription is always kept in History, even when pasting or saving the recording fails.

Recordings longer than about 13 minutes exceed 25 MB of audio. The Codex transcription endpoint publishes no size limit, and the public OpenAI transcription API rejects uploads above 25 MB, so Murmur warns before it uploads a recording of that size. The upload timeout grows with the recording length.

Meetings

Meetings are on by default and can be turned off in Settings → Advanced. Press Control+Option+Command+M, or choose Start meeting in the menu bar, to start recording; press it again to stop. You can pick other keys in Settings; if another app already holds them, the shortcut field says so. A red pill shows the elapsed time. A meeting stops by itself after 3 hours, with a warning 5 minutes before.

• Murmur records two tracks: your microphone ("Me") and the system audio, which holds the other participants. System audio needs macOS 14.4 or later on Apple silicon, and the System Audio Recording permission, which macOS asks for at the first meeting. Without it, meetings record the microphone only. When another app played sound during a meeting but its system audio holds nothing but silence while you spoke, the meeting says so and links to the privacy settings.
• After you stop, Codex transcribes the tracks in pieces of up to 10 minutes, at most 3 at a time. Processing resumes after a quit, without sending a transcribed piece again. Pieces Codex refuses are transcribed on the Mac when the local model is downloaded, and the meeting says so.
• On Apple silicon with macOS 15 or later, speaker detection runs on the Mac with FluidAudio. Its models are the local transcription download, fetched during the first meeting if they aren't there yet. Without them, or on macOS 14, where FluidAudio's speaker model crashes, each track keeps one speaker.
• Once the transcript is complete, Codex writes a title and a summary. The two tracks are then kept as one compressed audio.m4a (about 30 MB per hour), which you can delete while keeping the text.
• Murmur warns at the start of a meeting when less than 1.5 GB is free on the disk. Three hours take about 700 MB before compression.

Permissions after updating to the stable signature

Releases are signed with a stable Murmur certificate, so macOS keeps the Accessibility and microphone permissions from one update to the next. Earlier releases were signed ad hoc. After the first update from an ad hoc release, grant Accessibility and microphone access to Murmur once more. If the Accessibility switch already looks enabled but Murmur keeps waiting, use the repair button on the permissions screen, or follow the reset steps in BUILD.md.

Cancel stops a transcription before its result is inserted. If the text has already been pasted, cancellation still prevents a pending automatic Enter. The pasted text and its history entry remain available.

CLI

murmur --toggle-transcription
murmur --cancel
murmur --start-hidden
murmur --no-tray
murmur --debug
murmur --transcribe-file recording.wav --json

On Unix systems, SIGUSR2 also toggles transcription for a running instance:

pkill -USR2 -n murmur

Development

Prerequisites: the latest stable Rust toolchain and Bun.

bun install
bun run tauri dev
bun run build
bun run lint
bun run format:check
cargo test --manifest-path src-tauri/Cargo.toml

See BUILD.md for macOS dependencies, packaging, code signing, and notarization.

Architecture

Murmur uses Tauri 2 with a Rust backend and a React/TypeScript frontend. The runtime pipeline is:

microphone -> quiet-speech gain -> local (Parakeet) or Codex transcription -> dictionary -> AI cleanup (Codex) -> clipboard/paste

The backend owns audio capture, the Codex session and its refresh, transcription, AI cleanup, history, Insights, shortcuts, and macOS integration. The frontend owns onboarding, settings, history, Insights, and the recording overlay.

Privacy and security

In local mode your audio stays on this Mac. If ChatGPT is connected, the text (not the audio) is sent for AI cleanup, and audio goes to the cloud only when the local model can't help; History marks those. In cloud mode, audio sent for transcription leaves the computer and is handled by ChatGPT. Do not describe Murmur as offline while AI cleanup or the cloud fallback can run.

While AI cleanup is on, the transcript is sent to ChatGPT for cleanup. A cleanup request contains the transcript, your custom dictionary words, your dictation style, and the names and short excerpt around the cursor found in the frontmost window.

That context is read when a dictation starts, through Accessibility on the Mac and UI Automation on Windows, only while AI cleanup and the window-context setting are both on. It is never stored, and it leaves the computer only inside a cleanup request. Password fields and password managers are never read.

Meetings stay on the Mac, apart from two requests: the audio pieces sent to Codex for transcription, and the transcript sent to Codex for the summary. Speaker detection runs on the Mac. Its models are downloaded from Hugging Face once.

History, saved recordings, and Insights stay on the Mac. For Insights, Murmur records the frontmost app of each dictation and, in a browser, the site's host name. Insights store no dictated text. Murmur checks GitHub for updates, which can be turned off in Settings.

The Codex authentication cache is sensitive. Murmur reads the tokens needed for a request or a refresh, never logs them, and keeps no copy of its own. The only write is the token refresh described in Requirements.

Guarded paste saves the clipboard before it places the transcription there. By default, Murmur puts the saved items back, with every format they had, once the 8-second paste window ends. If you copy something else during that window, Murmur leaves your new clipboard alone. If insertion fails, Murmur keeps the transcription on the clipboard for manual paste. The Keep transcription option leaves the transcription on the clipboard after insertion.

Origin

Murmur is a fork of Handy, created by CJ Pais. It keeps parts of Handy's Tauri foundation, but most of Handy's original transcription engine and local-model stack have been removed. Murmur transcribes in the cloud through your ChatGPT account, or on the computer with Parakeet: FluidAudio on Apple silicon, and transcribe.cpp, the engine Handy uses, on Intel Macs and Windows.

Support

Murmur is free. To support development, buy a coffee.

License

MIT. See LICENSE. The original copyright notice is preserved as required by the license.