dpdfnet-ladspa

Packaging of the DPDFNet models (ceva-ip/DPDFNet) as LADSPA plugins for Linux. One .so per upstream model — users install only the variant they want.

Not a fork of DPDFNet: this repository is just the packaging layer (Rust + PKGBUILD + preparation scripts). Weights come from Hugging Face at build time; inference code uses OpenVINO Runtime.

Packaged models

Each entry becomes an independent Arch/BigLinux package (pkgname=dpdfnet-ladspa-<variant>) and a .so in /usr/lib/ladspa/.

Variant	Sample rate	Notes
baseline	16 kHz	Fastest, 0–8 kHz band
dpdfnet2	16 kHz	Upstream default, balanced quality/CPU
dpdfnet4	16 kHz	Deeper DPRNN stack
dpdfnet8	16 kHz	Highest quality in the 16 kHz range
dpdfnet2_48khz_hr	48 kHz	Full-band, balanced
dpdfnet8_48khz_hr	48 kHz	Full-band, highest quality

Layout

build.rs                   # model selection via DPDFNET_MODEL, emits model_const.rs
Cargo.toml / Cargo.lock    # cdylib crate, links to libopenvino_c via pkg-config
src/lib.rs                 # LADSPA implementation (STFT, OpenVINO inference, iSTFT)
scripts/prepare_model.py   # ONNX → OpenVINO IR + init_state extraction
pkgbuild/PKGBUILD          # split package, one package_*() per variant

Generated artefacts (gitignored): model/<name>/{model.xml,model.bin,init_state.bin}, target-<model>/, pkgbuild/{src,pkg}/, pkgbuild/*.onnx.

How the build works

1. prepare_model.py downloads/reads the upstream ONNX, converts it to OpenVINO FP32 IR (model.xml + model.bin) and extracts the initial recurrent state (init_state.bin).
2. build.rs reads model/<DPDFNET_MODEL>/* and generates Rust constants (sample rate, STFT geometry, state_size, LADSPA label, unique_id) embedded into the .so.
3. cargo build --release produces libdpdfnet_ladspa.so; the PKGBUILD repeats the cycle per variant, isolating CARGO_TARGET_DIR=target-<model>, and renames the output to libdpdfnet_<model>_ladspa.so at install time.

Result: each .so carries a single model, no runtime dispatch and no external weight files.

Manual build (without makepkg)

# 1. download upstream ONNX
mkdir -p /tmp/dpdfnet-models
curl -L -o /tmp/dpdfnet-models/dpdfnet2.onnx \
    https://huggingface.co/Ceva-IP/DPDFNet/resolve/main/onnx/dpdfnet2.onnx

# 2. convert to IR
python scripts/prepare_model.py --model dpdfnet2 \
    --src-onnx /tmp/dpdfnet-models/dpdfnet2.onnx

# 3. compile the .so for that model
DPDFNET_MODEL=dpdfnet2 CARGO_TARGET_DIR=target-dpdfnet2 \
    cargo build --release

Dependencies: rustc >= 1.85, cargo, openvino (runtime + headers), python with numpy, onnx, openvino.

Build via PKGBUILD (Arch/BigLinux/Manjaro)

cd pkgbuild
makepkg -si

Builds all six split packages. sha256sums is set to SKIP while the Hugging Face registry does not pin a revision.

Warning: Cargo.toml carries a [patch.crates-io] pointing at a patched ladspa crate in ../gtcrn-ladspa/ladspa/ladspa-patched. For reproducible builds outside the maintainer's environment, check out gtcrn-ladspa next to this repo, or drop the patch and use upstream ladspa = "0.3" if it already carries the needed fix.

Usage

Once any package is installed, the plugin shows up in LADSPA hosts (Ardour, Carla, EasyEffects via wrapper, etc.) under the model's label. For PipeWire filter-chain, reference libdpdfnet_<model>_ladspa.so and the label exposed by the plugin.

License

• Code in this repository: MIT (see Cargo.toml).
• DPDFNet models and ONNX weights: Apache-2.0 (ceva-ip/DPDFNet). The PKGBUILD installs a copy of the upstream LICENSE in /usr/share/licenses/<pkg>/ of each split, per Apache-2.0 §4(a).

Citation

Rika, Sapir, Gus (2025). DPDFNet: Boosting DeepFilterNet2 via Dual-Path RNN.