| 日本語 | English |
|---|
DeepFilterNet3 VST3 is a macOS audio plugin that embeds the official DeepFilterNet v0.5.6 model for real-time and offline noise reduction. It exports VST3 and CLAP bundles through nice-plug, accepts mono or stereo tracks, and keeps neural inference and sample-rate conversion on a persistent worker so the host audio callback remains nonblocking.
The compact custom editor provides Attenuation Limit and Mix controls with unit-separated numeric entry. No other controls or visualizations are included.
The plug-in bypassed and enabled:
- Features
- Preview
- Audio demo
- Tech stack
- Current validation scope
- Audio behavior
- Latency
- Requirements
- Build and model selection
- Install
- Usage
- Parameters
- Development and testing
- Project structure
- Troubleshooting
- Known limitations
- License
- Credits
- Official DeepFilterNet3 low-latency model by default, with the official standard model available as a separate build-time option.
- Mono and stereo input/output layouts with one mono inference stream.
- Arbitrary host block sizes through fixed, timestamped worker chunks.
- Streaming conversion for 44.1, 48, 88.2, 96, 176.4, and 192 kHz host rates.
- Reported latency with sample-aligned dry/wet mixing.
- The same DSP, resamplers, timeline, and reset protocol in Realtime, Buffered, and Offline modes.
- Lock-free callback transport and a latency-aligned dry fallback when a worker result is late.
- Compact English custom editor with only Attenuation Limit and Mix sliders.
- VST3 and CLAP exports with stable plugin and parameter IDs.
| Component | Role |
|---|---|
| Rust 2021 workspace | Plugin, DSP bridge, tests, and bundle task |
| nice-plug 0.2.3 | VST3/CLAP framework and exports |
| nice-plug-egui 0.3.0 / egui 0.35.0 | Embedded two-slider custom editor |
| DeepFilterNet 0.5.6 | Official embedded model and Tract inference |
| rubato 0.14.1 | Persistent fixed-size sample-rate conversion |
| rtrb 0.3.3 | Lock-free worker queues |
The current implementation is built and tested on Apple Silicon with macOS
26. Automated validation includes 25 Rust tests and pluginval strictness 5
with callback allocation assertions. pluginval opened the custom editor both
idle and during processing, exercised editor automation plus 44.1, 48, and 96
kHz processing, and completed with SUCCESS.
A user-confirmed test in DaVinci Resolve 21 completed a successful Deliver export with an earlier bundle. That bundle predated the custom editor, so the result is not UI validation for the current build. The broader repeatability, interaction, latency, and multi-rate Resolve smoke-test matrix has not yet been completed. Windows, Linux, and Intel macOS builds have not been validated.
The embedded model always receives one channel:
- Mono input is passed directly to inference.
- Stereo input is downmixed as
(left + right) / 2for inference. - The mono wet result is copied to both stereo outputs.
- Each stereo channel retains its own dry signal before the aligned dry/wet mix.
The plugin delays both dry and wet output to the reported latency. During startup or a real-time worker underrun, the affected samples use dry audio from the same delayed timestamp instead of silence or a stale wet frame. Offline mode uses the same worker pipeline and may wait up to two seconds for the required timestamped result.
Unsupported sample-rate or host-buffer geometry, model startup failure, and other initialization failures select unchanged direct bypass with zero reported latency.
Latency is calculated from live model metadata, both resamplers, and two host quanta reserved for nonblocking collection and inference. The official low-latency model reports a 48 kHz FFT size of 960, hop size of 480, zero lookahead, and 480 samples of intrinsic model delay.
| Host rate | Host quantum | Reported latency | Impulse validation |
|---|---|---|---|
| 44.1 kHz | 441 samples | 1,764 samples (40 ms) | Within 1 sample |
| 48 kHz | 480 samples | 1,440 samples (30 ms) | Exact |
| 96 kHz | 960 samples | 3,840 samples (40 ms) | Within 1 sample |
Mix values of 0%, 50%, and 100% remain peak-aligned at the reported latency. The other declared rates use the same checked formula and streaming converter geometry.
- Apple Silicon Mac running macOS 26.x or later for the validated configuration.
- Rust 1.87 or later to build nice-plug 0.2.3.
- A VST3- or CLAP-compatible host.
The build downloads Rust dependencies and the pinned official DeepFilterNet v0.5.6 source/model archive.
Clone the repository and build the default low-latency model:
git clone https://github.com/Shuichi346/DeepFilterNet3-VST3.git
cd DeepFilterNet3-VST3
cargo xtask bundle deepfilter-vst --releaseGenerated bundles:
target/bundled/deepfilter-vst.vst3
target/bundled/deepfilter-vst.clap
To build the official standard model instead of the default low-latency model:
cargo xtask bundle deepfilter-vst --release --no-default-features --features model-standardThe model features are mutually exclusive. Exactly one of model-ll or model-standard must be enabled.
For a user-only VST3 installation on macOS:
mkdir -p "$HOME/Library/Audio/Plug-Ins/VST3"
cp -R target/bundled/deepfilter-vst.vst3 "$HOME/Library/Audio/Plug-Ins/VST3/"For CLAP hosts:
mkdir -p "$HOME/Library/Audio/Plug-Ins/CLAP"
cp -R target/bundled/deepfilter-vst.clap "$HOME/Library/Audio/Plug-Ins/CLAP/"Restart or rescan the host after installation. Local builds are not distributed with a Developer ID signature or Apple notarization.
- Build and install the VST3 or CLAP bundle, then restart or rescan the host.
- Add DeepFilter Noise Reduction to a mono or stereo audio track.
- Open the plug-in editor. It contains only Attenuation Limit and Mix sliders.
- Leave Mix at 100% for the fully enhanced signal, or reduce it to blend in the latency-aligned dry channel.
- Adjust Attenuation Limit to cap the amount of noise attenuation. A 0 dB setting selects aligned raw audio while keeping model state advancing.
The host receives the plugin's calculated latency during initialization. If the requested host configuration is unsupported, the plugin remains available but passes audio through unchanged and reports zero latency.
The compact custom editor and host-generated parameter panels both control the same two automatable parameters. Host automation and external parameter changes remain synchronized with the sliders.
| Parameter | Range | Default | Behavior |
|---|---|---|---|
| Attenuation Limit | 0–100 dB | 100 dB | Limits the attenuation applied by DeepFilterNet. At effectively 0 dB, the model still advances while the aligned raw path is selected. |
| Mix | 0–100% | 100% | Blends latency-aligned per-channel dry audio with the mono wet result. |
Build a debug bundle with nice-plug's callback allocation assertions:
cargo xtask bundle deepfilter-vst --features nice-plug/assert_process_allocsRun the bounded library and plugin validation gate:
cargo test -p deepfilter-vst --lib && \
/Applications/pluginval.app/Contents/MacOS/pluginval \
--strictness-level 5 \
--validate-in-process \
target/bundled/deepfilter-vst.vst3The VST3 bundle used for pluginval should be the allocation-asserting debug artifact from the preceding command.
Create the Apple Silicon release package after building the release bundles:
cargo xtask bundle deepfilter-vst --release
./scripts/package-release.shThe script reads the version from plugin/Cargo.toml. You can also pass an explicit version:
./scripts/package-release.sh 0.5.0It verifies that both bundles are thin arm64 binaries with valid ad-hoc signatures, then creates:
dist/DeepFilterNR-v0.5.0-macos-arm64.zip
dist/DeepFilterNR-v0.5.0-macos-arm64.zip.sha256
The ZIP contains both plug-in bundles, installation instructions, required license notices, and checksums. Existing packages are never overwritten. The script does not install or publish anything.
plugin/src/lib.rs Plugin metadata, lifecycle, host layouts, and exports
plugin/src/params.rs Attenuation Limit and Mix parameters
plugin/src/editor.rs Fixed-size English two-slider custom editor
plugin/src/bridge.rs Callback-side buffering, alignment, and fallback
plugin/src/dsp.rs Worker DSP core and latency calculation
plugin/src/model.rs DeepFilterNet model wrapper and metadata
plugin/src/resampler.rs Checked persistent sample-rate conversion
plugin/src/worker.rs Worker lifecycle, queues, reset, and status
xtask/ VST3/CLAP bundle command
scripts/ Release packaging tools
PLANS.md records the implementation and validation evidence. CHANGELOG.md
and NOTES.md record release and maintenance information.
If a host keeps discovering an older local build, clean and recreate the release bundle before reinstalling it:
cargo clean
cargo xtask bundle deepfilter-vst --releaseConfirm that the VST3 or CLAP directory matches the installation paths above, then restart or rescan the host. A locally built bundle is not Developer ID signed or notarized, so macOS host security behavior may differ from a distributed signed plugin.
- The wet path is mono by design; stereo spatial differences remain only in the dry contribution.
- Real-time scheduling delays can temporarily substitute aligned dry audio for missing enhanced output.
- Worker/model startup is bounded to ten seconds; a startup failure selects direct bypass.
- Unsupported rate or buffer configurations select direct bypass rather than resampling approximately.
- A successful Deliver export has been user-confirmed in DaVinci Resolve 21, but the full repeatability, interaction, latency, and multi-rate host matrix remains unverified.
- Only the Apple Silicon macOS configuration described above has been validated.
MIT License. Required notices for redistributed components are in Third-Party Notices.
- DeepFilterNet by Hendrik Schröter and contributors.
- nice-plug by RustAudio contributors.
- rubato for streaming sample-rate conversion.