Skip to content
nu-syncPublic

About

Local Nemotron dictation in Rust

Resources

Contributing

Stars

0 stars

Watchers

0 watching

Forks

Repository files navigation

Nemo

Local voice dictation for Apple Silicon Macs. Tap right Option to record, then tap it again to transcribe and paste into the focused app.

Audio is processed on your Mac. No account or API key is required.

After installation and model download, dictation works offline. The controller and inference runtime deny external networking; a constrained helper fetches only the pinned model. See the local-only audit for controls, license checks, verification, and the separate network requirements of source installation.

Engines

Nemo runs one of two local inference engines, chosen in config.toml without a rebuild:

backend What it is Needs Memory while loaded
nemo-speech The default: the bundled NeMo-Speech.cpp runtime. Runs after a standard install. Nothing extra about 20 GB
mlx-audio Opt-in: MLX-Audio with speculative decoding. 1.50–1.65x faster on 5–100 second dictations, same word errors. uv, Python 3.12+, 21.6 GB of MLX weights about 27 GB

nemo-speech is the default. mlx-audio is opt-in: set it up with nemo backend setup mlx-audio, then select it with nemo backend use mlx-audio. If it fails to load, Nemo falls back to nemo-speech. Numbers are from the November 9 engine comparison on an M5 Max.

Install

You need macOS 26 or newer, an Apple Silicon Mac, Rust, and Homebrew. Allow about 22 GB for either model download, plus space for build files.

Install Apple’s Command Line Tools if you haven’t already:

xcode-select --install

Finish that installer, then open Terminal and run:

cargo install --locked --git https://github.com/nu-sync/nemo.git nemo-dictate

Choose one of the following setup paths before opening Nemo.

Default engine

nemo

The first nemo run installs the build dependencies through Homebrew, builds the inference runtime, and creates ~/Applications/Nemo.app. Nemo appears as a dot in the menu bar and automatically downloads the Nemotron model. Open its menu to check the download.

Setup happens once. Later, run nemo or open Nemo from ~/Applications.

MLX-Audio only

Install uv and make sure python3 --version reports 3.12 or newer. Run the following before opening Nemo for the first time:

nemo install
uv sync --locked --script "$HOME/Applications/Nemo.app/Contents/Resources/runtime/mlx-serve.py"
nemo backend setup mlx-audio
nemo backend use mlx-audio

Open ~/Library/Application Support/Nemo/config.toml and replace the existing model line with:

model = "/dev/null"

This setting skips the automatic download of the default GGUF weights. MLX-Audio loads the weights downloaded by nemo backend setup mlx-audio.

open -a "$HOME/Applications/Nemo.app"
nemo backend list

Look for * mlx-audio: weights ready. The app bundle includes both runtimes. On a fresh install with these steps, the list shows the default weights as not provisioned. The menu bar may report Transcription blocked: model missing because its readiness check looks for GGUF weights. Use nemo backend list to check the MLX weights.

If you later switch to nemo-speech, remove the model line to restore its default download path.

Dictate

Wait for the model download to finish, then:

  1. Enable Nemo in System Settings → Privacy & Security → Accessibility. This enables right Option and automatic paste.
  2. Click where you want the text to go.
  3. Tap right Option. Allow Microphone access when prompted. Speak when Recording appears at the bottom of the screen.
  4. Tap right Option again. The indicator changes to Transcribing, then disappears when Nemo finishes copying and pasting your words.

Keep the destination app focused until paste finishes. Cancel in Nemo’s menu discards the current operation. Quit closes Nemo and releases the model.

The menu bar dot is gray when idle, red when recording, and amber during setup or transcription. Left Option and Option used with another key keep their normal behavior.

The indicator appears above the Dock on the display containing your pointer. It lets clicks pass through and keeps your destination app focused.

Settings

Choose Settings… from Nemo’s menu, or run nemo settings.

Setting What it does
Remove trailing period Removes the final . from each transcript. Off by default.
Launch on startup Opens Nemo when you log in. macOS may ask you to approve the login item.
Sound pack Choose Handy Marimba or macOS Recording. Preview the start and stop sounds.

These settings save immediately. For other options, choose Open configuration from the menu. The file is ~/Library/Application Support/Nemo/config.toml; restart Nemo after editing it.

The model loads while you record and unloads after 60 idle seconds by default. To keep it ready for five minutes, change this value in your configuration:

idle_unload_seconds = 300

See advanced configuration for shortcuts, recording limits, custom models, and inference tuning.

Update

Finish your current dictation, then run:

cargo install --locked --force --git https://github.com/nu-sync/nemo.git nemo-dictate
nemo

Nemo keeps your model and settings and retains the previous app in ~/Applications/.Nemo.backups/. It refuses to replace an app that is recording or transcribing.

If an update changes the app’s signing identity, the installer stops before replacing it. To accept that update:

nemo install --allow-permission-reset
nemo

You may need to grant Microphone and Accessibility access again.

Troubleshooting

nemo doctor
nemo status

doctor reports model and runtime paths, the input device, and permissions for both the command and the running app. Logs are in ~/Library/Application Support/Nemo/daemon.log and inference.log.

  • No recording or paste: check Nemo’s Microphone and Accessibility entries in System Settings.
  • Interrupted model download: choose Download model in the menu, or quit Nemo and run nemo again. Partial downloads are retained.
  • Build failed: fix the reported dependency or build error and run nemo again. Completed build files are retained.

Development

Build and install the current checkout with Rust:

cargo xtask check
cargo xtask install

The update command retains the previous app and refuses to interrupt dictation. Python and Bash are not required for building or installing Nemo.

See development and advanced configuration for source builds, CLI commands, packaging, tests, and the application design. See contributing and release preparation for development and publication requirements. The project’s verify-nemo skill exercises real speaker playback, microphone capture, transcription, paste, and idle unloading.

Licenses

Nemo is MIT licensed. It uses NVIDIA Nemotron 3.5 Transcribe, the AtomicChat IQ4_NL conversion, and NeMo-Speech.cpp. Model, runtime, and sound licenses are included in Nemo.app/Contents/Resources/licenses.

About

Local Nemotron dictation in Rust

Resources

Contributing

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages