Linux Broadcast

NVIDIA Broadcast-style voice processing for Linux, powered by NVIDIA AFX.

Linux Broadcast is a native NVIDIA Audio Effects SDK (AFX) voice processor for PipeWire with a minimal Tauri desktop interface. The codebase supports NVIDIA's Noise Removal, experimental BNR 2.0, Room Echo Removal, combined Noise + Room Echo, and Studio Voice Low Latency models. There are no substitute denoising backends.

Features

NVIDIA Noise Removal and experimental BNR 2.0
Room Echo Removal and combined Noise + Room Echo processing
Studio Voice Low Latency
Automatic RTX 20, 30, 40, and 50-series model selection
A persistent Linux Broadcast Microphone for PipeWire applications
Physical microphone and monitoring-output selection
Easy Effects-compatible routing
Per-effect intensity, VAD, and frame-size controls
Background operation, system tray, and start-at-login support

Installation

Linux Broadcast requires an RTX GPU, the proprietary NVIDIA driver, PipeWire, and WebKitGTK 4.1. Complete generation-specific RPM, DEB, and portable bundles are available from the download homepage. Each bundle contains the app, native plugin, NVIDIA AFX runtime, CUDA, TensorRT, Noise Removal, BNR 2.0, Room Echo, combined Noise + Room Echo, and low-latency Studio Voice. Users do not need an NGC login or a separate CUDA toolkit after downloading the correct package.

Choose the release matching the installed GPU generation:

Download	GPUs
`linux-broadcast-0.2.0-rtx20-x86_64.tar`	GeForce RTX 20 and Quadro RTX
`linux-broadcast-0.2.0-rtx30-x86_64.tar`	GeForce RTX 30 and RTX A-series
`linux-broadcast-0.2.0-rtx40-x86_64.tar`	GeForce RTX 40 and RTX Ada
`linux-broadcast-0.2.0-rtx50-x86_64.tar`	GeForce RTX 50 and RTX PRO Blackwell

Extract the matching archive and run its launcher:

tar -xf linux-broadcast-0.2.0-rtx40-x86_64.tar
cd linux-broadcast-0.2.0-rtx40-x86_64
./linux-broadcast

Verify the download with the accompanying SHA256SUMS file. Portable bundles are not registered with the system package manager. RPM and DEB builds remain available for maintainers using the packaging instructions below.

This is a public preview. The RTX 40 bundle is hardware-probed on a GeForce RTX 4080. RTX 20, 30, and 50 bundles pass the same architecture, model, and runtime checks but still need compatibility reports from those GPUs.

Fedora, openSUSE, and compatible RPM distributions

sudo dnf install ./linux-broadcast-<version>-rtx<series>.x86_64.rpm

Remove it with sudo dnf remove linux-broadcast.

Debian, Ubuntu, and compatible distributions

sudo apt install ./linux-broadcast_<version>_rtx<series>_amd64.deb

Remove it with sudo apt remove linux-broadcast.

RPM and DEB installations are available from the desktop menu or as linux-broadcast. Portable bundles are launched from their extracted directory. Source builds use AFX_SDK_ROOT or ~/.local/share/linux-broadcast/nvidia/current.

GPU and model selection

The application selects an architecture-specific TensorRT model from the GPU's compute capability:

GeForce generation	Compute capability	Model directory	NVIDIA package equivalent
RTX 20 / Turing	7.5	`sm_75`	T4
RTX 30 / Ampere	8.6	`sm_86`	A10
RTX 40 / Ada	8.9	`sm_89`	L40
RTX 50 / Blackwell	12.0	`sm_120`	RTX PRO 6000

The package equivalents identify matching compute architectures. They do not imply official GeForce support from NVIDIA.

This covers desktop and laptop GeForce RTX cards, Quadro RTX, RTX A-series, RTX Ada, and RTX PRO Blackwell cards. On multi-GPU systems the app prefers a compatible RTX GPU with an installed model. Set LINUX_BROADCAST_GPU_INDEX to override automatic selection.

./scripts/detect-gpu.sh

The Linux SDK's supported-device selector rejects GeForce cards, so the engine retains CUDA device 0 and loads the matching model directly.

Licensed AFX installation

Set AFX_SDK_ROOT to an extracted AFX 2.x SDK. NVIDIA files are never committed to Git. Release packages stage the permitted runtime subset and applicable license notices directly from this directory.

$AFX_SDK_ROOT/
├── nvafx/include/nvAudioEffects.h
├── nvafx/lib/libnv_audiofx.so
└── features/
    ├── denoiser/
    ├── dereverb/
    ├── dereverb_denoiser/
    └── studio_voice/

With NVIDIA AI Enterprise access, install the 48 kHz feature packages for the detected RTX architecture:

export AFX_SDK_ROOT=/path/to/Audio_Effects_SDK
./scripts/install-licensed-denoiser.sh

The installer reads the standard NGC CLI credential store when NGC_API_KEY is not set. When run in a terminal without either configuration, it prompts for a personal key without echoing or saving it.

To create an NGC personal key:

Sign in to NVIDIA NGC.
Select Generate Personal Key, give it a descriptive name, and grant the catalog access required by the AFX resources in your organization.
Copy the key when NVIDIA displays it. NVIDIA does not display it again.
Run ./scripts/install-licensed-denoiser.sh and paste the key at the secure prompt, or configure it with ngc config set.

Release RPM and DEB files already contain their targeted models and do not ask for an NGC key during package installation. The key flow is for source builds, custom architecture packages, and future models that are not in a release.

Native engine

The engine configures AFX for 48 kHz mono and one stream. Regular effects can use NVIDIA-supported 10 ms or 20 ms frames; Studio Voice Low Latency is fixed to 10 ms. VAD, intensity, and effect-version parameters are applied only to models that support them.

cmake -S native -B build/native -DAFX_SDK_ROOT="$AFX_SDK_ROOT"
cmake --build build/native -j
export LD_LIBRARY_PATH="$AFX_SDK_ROOT/nvafx/lib:$AFX_SDK_ROOT/features/denoiser/lib:$AFX_SDK_ROOT/features/dereverb/lib:$AFX_SDK_ROOT/features/dereverb_denoiser/lib:$AFX_SDK_ROOT/features/studio_voice/lib:$AFX_SDK_ROOT/external/cuda/lib:${LD_LIBRARY_PATH:-}"
./build/native/linux-broadcast-afx --probe noise
./build/native/linux-broadcast-afx --probe bnr2
./build/native/linux-broadcast-afx --probe room_echo
./build/native/linux-broadcast-afx --probe noise_room_echo
./build/native/linux-broadcast-afx --probe studio_voice

The probe loads the selected model and processes one frame.

Tauri UI

ui/ is a working Tauri 2 desktop wrapper around the native AFX engine. Its Rust bridge detects physical PipeWire microphones, starts a persistent pw-cli controller, loads the selected AFX effect chain, and publishes the processed stream as linux_broadcast.source. Stopping the effect or quitting from the tray removes the virtual microphone. Closing the window keeps it active when Run in background is enabled.

The UI exposes the input source, installed effect modes, intensity, headphone monitoring, GPU architecture, and model/plugin readiness. Monitoring is off by default to prevent speaker feedback. Easy Effects is used as the monitoring source when available.

SDK-specific VAD and frame sizing live in the collapsed Advanced view. A prebuilt installation starts with Noise + Room Echo, 70% intensity, VAD off, and 20 ms frames. Every installed effect remembers its own tuning.

The Settings page contains the two lifecycle controls:

Run in background closes the window to the tray while keeping NVIDIA AFX and the virtual microphone active.
Start at login installs and enables a user-level systemd service. It launches the app hidden with the saved microphone, effect, and tuning. Enabling it also enables background mode.

The release build can also be installed or removed from the terminal:

./scripts/install-background-service.sh
./scripts/uninstall-background-service.sh

The service installs only into the current user's ~/.local and ~/.config/systemd/user directories; it does not require root.

Build from source

The source build requires CMake 3.20 or newer, a C++20 compiler, LADSPA headers, Node.js 20.19 or newer, npm, the stable Rust toolchain, and the Linux dependencies required by Tauri 2.

Fedora:

sudo dnf install gcc-c++ cmake ladspa-devel nodejs npm rust cargo \
  pkgconf-pkg-config webkit2gtk4.1-devel openssl-devel \
  libappindicator-gtk3-devel librsvg2-devel libxdo-devel pipewire-utils

Debian or Ubuntu:

sudo apt update
sudo apt install build-essential cmake ladspa-sdk pkg-config curl wget file \
  libwebkit2gtk-4.1-dev libssl-dev libayatana-appindicator3-dev \
  librsvg2-dev libxdo-dev pipewire-bin

Install a supported Node.js release and the stable Rust toolchain if the distribution versions do not meet the requirements above. Then clone the repository, configure the licensed SDK, and build the native plugin followed by the desktop application:

git clone https://github.com/arzvaak/linux-broadcast.git
cd linux-broadcast

export AFX_SDK_ROOT="$HOME/.local/share/linux-broadcast/nvidia/current"
cmake -S native -B build/native-cmake \
  -DAFX_SDK_ROOT="$AFX_SDK_ROOT" \
  -DBUILD_TESTING=ON
cmake --build build/native-cmake --parallel
ctest --test-dir build/native-cmake --output-on-failure

npm ci --prefix ui
npm run build --prefix ui
cargo test --manifest-path ui/src-tauri/Cargo.toml --locked
npm run tauri --prefix ui -- build --no-bundle
./ui/src-tauri/target/release/linux-broadcast

The application automatically uses the plugin at build/native-cmake/liblinux_broadcast_afx_ladspa.so. AFX_SDK_ROOT and LINUX_BROADCAST_PLUGIN override the default SDK and plugin paths.

Build RPM and DEB packages

Each release package contains one GPU generation so downloads do not carry models that cannot run on the target machine. Every package contains all 48 kHz models used by Linux Broadcast and the measured CUDA, TensorRT, and NVIDIA AFX runtime closure needed by those models. Unused cuDNN and Studio Voice HQ files are deliberately excluded. NVIDIA files remain outside Git and are read from AFX_SDK_ROOT only while packaging.

Build the RPM and DEB for a generation after installing its model variant into the SDK tree:

export AFX_SDK_ROOT=/path/to/Audio_Effects_SDK
./scripts/build-packages.sh rtx40

Supported release targets are:

Release target	GPUs	Model architecture
`rtx20`	RTX 20 and Quadro RTX	`sm_75`
`rtx30`	RTX 30 and RTX A-series	`sm_86`
`rtx40`	RTX 40 and RTX Ada	`sm_89`
`rtx50`	RTX 50 and RTX PRO Blackwell	`sm_120`

The resulting files are written to:

build/releases/<version>/<series>/linux-broadcast-<version>-<series>.x86_64.rpm
build/releases/<version>/<series>/linux-broadcast_<version>_<series>_amd64.deb

The staging script copies the five model files selected by the application for the target architecture, the measured runtime closure, and applicable license notices. It rejects cuDNN, Studio Voice HQ, and credential-like files. NGC configuration and credentials are never read into the package.

Create the four uncompressed portable release archives with:

export AFX_SDK_ROOT=/path/to/Audio_Effects_SDK
./scripts/build-portable-releases.sh

Build the complete twelve-artifact matrix against the Ubuntu 22.04 compatibility baseline with Podman, then publish the verified files to the configured Netcup host:

export AFX_SDK_ROOT=/path/to/Audio_Effects_SDK
./scripts/build-jammy-releases.sh
./scripts/publish-netcup-release.sh

verify-release-artifact.sh checks the selected GPU architecture, five models, runtime closure, native plugin, application, and NVIDIA license notices in each RPM, DEB, or portable archive before publication.

Primary references

About me

I'm Ayush, also known as Arzvak. I study electrical and electronics engineering and build software around the problems I run into. I started Linux Broadcast because Linux users with RTX hardware should have a native, polished voice-processing tool instead of giving up the NVIDIA hardware they already own.

Contributing

Bug reports, compatibility results, documentation improvements, and focused pull requests are welcome. Read CONTRIBUTING.md before making a change. Please report security issues through the process in SECURITY.md, not a public issue.

License

The Linux Broadcast source code is available under the MIT License. NVIDIA AFX libraries and models remain NVIDIA software and are distributed in release packages under their accompanying NVIDIA license terms. They are not part of the MIT-licensed source repository.

Linux Broadcast is an independent project and is not sponsored or endorsed by NVIDIA Corporation. NVIDIA, RTX, and NVIDIA Broadcast are trademarks of NVIDIA Corporation.

Acknowledgements

NVIDIA Maxine Audio Effects SDK
PipeWire for the Linux audio graph
Tauri for the desktop application framework
BluCast for demonstrating another community-built Maxine workflow on Linux

Name		Name	Last commit message	Last commit date
.github/workflows		.github/workflows
docs		docs
native		native
ops/netcup		ops/netcup
packaging		packaging
scripts		scripts
systemd		systemd
ui		ui
.gitignore		.gitignore
CHANGELOG.md		CHANGELOG.md
CONTRIBUTING.md		CONTRIBUTING.md
LICENSE		LICENSE
README.md		README.md
SECURITY.md		SECURITY.md

Navigation Menu

Repository files navigation

Linux Broadcast

Features

Installation

Fedora, openSUSE, and compatible RPM distributions

Debian, Ubuntu, and compatible distributions

GPU and model selection

Licensed AFX installation

Native engine

Tauri UI

Build from source

Build RPM and DEB packages

Primary references

About me

Contributing

License

Acknowledgements

About

Releases

Packages

Used by

Contributors

Languages

Latest commit

History

Folders and files

Repository files navigation

Linux Broadcast

Features

Installation

Fedora, openSUSE, and compatible RPM distributions

Debian, Ubuntu, and compatible distributions

GPU and model selection

Licensed AFX installation

Native engine

Tauri UI

Build from source

Build RPM and DEB packages

Primary references

About me

Contributing

License

Acknowledgements

About

Topics

Resources

Contributing

Security policy

Stars

Watchers

Forks

Releases

Packages

Used by

Contributors

Languages