OpenSuperMLX

Private, real-time dictation for your Mac.
Press a shortcut in any app, speak, and your words appear where you're typing.
Transcribed on your Mac, never uploaded.

brew tap axot/tap && brew install --cask opensupermlx
or download the DMG from the latest release

How it works

1. In any app, press ⌥ + `, or hold it while you talk.
2. Speak. The text appears live as you go.
3. Stop, and the transcript is pasted where your cursor is.

Why OpenSuperMLX

• Private by design. Speech recognition runs entirely on your Mac, using Qwen3-ASR on MLX and the Neural Engine. There's no account and no subscription, and your audio is never sent anywhere.
• Works wherever you type. Slack, email, documents, your editor, an AI chat box: one global shortcut, and the text lands in the app you're using.
• Real time, even in long sessions. Words stream in while you talk, and long meetings keep transcribing.
• Made for multilingual speakers. It detects the language automatically across 19 languages, fixes spacing and punctuation for Chinese, Japanese, and Korean, and writes numbers as digits in Chinese and English.
• Recordings you can come back to. Every session is kept, with its audio, in a searchable history. If a save fails, you can retry instead of losing the take. You can also drop in an existing audio file to transcribe it.
• Optional LLM cleanup. A second shortcut sends only the text to your own AWS Bedrock or OpenAI-compatible model to fix misheard words and punctuation.

Made for

• Writing faster: meeting notes, chat replies, prompts, documents, and follow-ups.
• Meetings and calls: with headphones on, OpenSuperMLX can record your microphone and the other side of the call together.
• Developers and AI agents: a local MCP server lets agents list, search, and follow live transcripts, and a CLI covers transcription, benchmarks, and diagnostics.

Features

Dictation

• Tap the shortcut to start and stop, or hold it to record only while it's pressed. Press Escape to cancel.
• Live streaming transcription.
• Switch between built-in, USB, Bluetooth, and Continuity microphones.
• Optional system-audio capture with headphones. On speakers it records the microphone only, to avoid echo.

Text

• Automatic language detection, or pick one of 19 languages.
• Asian autocorrect for Chinese, Japanese, and Korean.
• Numbers, dates, and amounts written as digits (inverse text normalization) for Chinese and English.
• Optional LLM correction through AWS Bedrock or any OpenAI-compatible API. Long transcripts are split to fit the model's limits.

History and stats

• Searchable transcript history. Regenerate any transcript from its audio.
• Recordings stored as 16 kHz mono AAC at 48 kbps. Imported audio keeps its original format.
• Drag and drop audio files to transcribe them in a queue.
• A stats dashboard with sessions, streaks, spoken time, and time saved compared with typing.

Reliability

• If a recording can't be saved, a recovery panel lets you retry, copy the transcript, or discard it.
• Signed with an Apple Developer ID and notarized by Apple.
• Automatic updates that wait until you've finished recording.

For developers

• A headless CLI: transcribe, stream-simulate, correct, config, recordings, queue, mic, model, benchmark, and diagnose.
• An opt-in Transcript MCP server on 127.0.0.1 for agent access to live transcripts.

Install

Homebrew

brew tap axot/tap
brew install --cask opensupermlx

Download

Download the DMG from the latest release, open it, and drag OpenSuperMLX to Applications.

Permissions

On first launch, macOS asks for microphone and accessibility access. OpenSuperMLX needs both: the microphone to record, and accessibility to paste text into other apps. Capturing system audio also needs Screen Recording permission.

Updates

OpenSuperMLX checks GitHub releases once a day and shows an update window when a new version is available. The window waits until you finish recording and stop typing, and the app restarts to install only when nothing is being recorded, saved, or transcribed. Choose Check for Updates… in the menu bar menu to check now, or turn off Settings → Advanced → Check for updates automatically.

Version 0.1.1 and earlier can't update themselves: install the next release once from the releases page or with brew upgrade --cask opensupermlx. Homebrew installs keep working with brew upgrade as before.

Requirements

• macOS 15 or later
• A Mac with Apple Silicon

Model

OpenSuperMLX uses Qwen3-ASR-1.7B-5bit (about 1.8 GB), downloaded from Hugging Face during first-launch setup.

Settings → Model → Neural Engine audio encoder (on by default) runs the model's audio encoder on the Neural Engine instead of the GPU. The first time the model loads, the app downloads a Core ML version of the encoder (about 300 MB). Turn the setting off to use the GPU encoder instead.

Shortcuts

Shortcut	Action
Tap ⌥ + `	Start or stop recording
Hold ⌥ + `	Record only while held
Tap ⌥ + ⇧ + `	Start or stop recording with LLM correction
Escape	Cancel the active recording

Change them in Settings → Shortcuts.

CLI

The app binary doubles as a headless CLI:

BINARY=build/Build/Products/Debug/OpenSuperMLX.app/Contents/MacOS/OpenSuperMLX
$BINARY diagnose --json
$BINARY help transcribe

See docs/cli.md for the full command reference.

Build from source

git clone git@github.com:axot/OpenSuperMLX.git
cd OpenSuperMLX
git submodule update --init --recursive
brew install cmake libomp rust ruby
gem install xcpretty
./run.sh build

After the first build, Swift-only changes can use the faster incremental xcodebuild command in AGENTS.md. ./run.sh rebuilds every native component and launches the app, and ./run.sh build builds without launching.

Support

If something goes wrong:

1. Search the existing GitHub issues.
2. Run diagnose --json from the CLI and collect the relevant logs as described in docs/logging.md.
3. Open a new issue with the steps to reproduce it and the diagnostic output.

If a Recording Not Saved panel appears, free up disk space or fix the storage problem, then choose Retry. Copy Transcript puts the text on the clipboard. Cancel Save permanently discards the audio and transcript, after asking you to confirm. See docs/audio-diagnostics.md for details on recording analysis and recovery.

Acknowledgments

OpenSuperMLX is forked from OpenSuperWhisper by @Starmel. Thanks to the original project for the foundation.

License

OpenSuperMLX is released under the MIT License. See LICENSE for details.