speech-enhancement-native

Native Rust inference engines for published speech-enhancement models, built as LADSPA and PipeWire plugins for BigLinux.

The models come from their authors: DeepFilterNet3 (Schröter et al.), DPDFNet (Ceva), GTCRN-AEC (LocalVQE), Silero VAD (Silero Team) and FireRedVAD (FireRedTeam). This repository reimplements their inference path without ONNX Runtime, OpenVINO or PyTorch. Recurrent weights are int8 with int16 activations (W8A16), and the kernels are chosen at runtime for the CPU: scalar, SSE4.1, AVX or AVX2+FMA. The int8 matrix products accumulate exactly in integers, so their result does not depend on the tier.

Packages

Package	Plugin	File	Label / name
deepfilternet3-native	DeepFilterNet3 noise suppression (default)	/usr/lib/ladspa/libdfn3_ladspa.so	deep_filter_net3_rs_mono
	DeepFilterNet3-LL, larger model, no lookahead	/usr/lib/ladspa/libdfn3ll_ladspa.so	deep_filter_net3_ll_rs_mono
	Lookahead peak limiter	/usr/lib/ladspa/liblimiter_ladspa.so	biglinux_lookahead_limiter_mono
dpdfnet-native	DPDFNet-2 48 kHz HR noise suppression	/usr/lib/ladspa/libdpdfnet_native.so	dpdfnet_native_48hr
	The same model over whole files, on several threads (how)	/usr/bin/dpdfnet-enhance
gtcrn-aec-native	GTCRN-AEC echo cancellation	/usr/lib/spa-0.2/aec/libspa-aec-gtcrn.so	library.name = aec/libspa-aec-gtcrn

All plugins take mono at 48 kHz (the limiter at any rate) and accept any host block size. The DeepFilterNet3 and DPDFNet plugins include a voice gate that mutes the output while nobody speaks, with Silero or, by choice, FireRedVAD as its detector, at one of five sensitivity levels. It raises the latency of every engine to 66 ms, from 40 ms for DeepFilterNet3, 20 ms for DeepFilterNet3-LL and 60 ms for DPDFNet (details). It is on by default in the DeepFilterNet3 plugins and off in DPDFNet, where it costs nothing until a host turns it on (controls). Each LADSPA plugin reports its current delay, in samples, on an output control port named latency, which PipeWire's filter-chain adds to the node latency.

CPU cost

CPU seconds per minute of audio, one core, measured with perf stat (method):

Engine	i5-13400 (AVX2)	i3-2375M (AVX, 2011)
DeepFilterNet3	0.8	5.2
DeepFilterNet3-LL	2.1	15.2
DPDFNet-2 48 kHz HR	3.9	27.3
GTCRN-AEC (16 kHz core)	1.2	7.2
Voice gate with Silero, added to any engine	0.33	not measured
Voice gate with FireRedVAD, added to any engine	0.55	not measured

Build

Rust 1.98.1 or newer (rust-toolchain.toml).

cargo build --release --locked --workspace --exclude dpdfnet-native
cargo build --locked --profile release-unwind -p dpdfnet-native

cargo test --release --locked --workspace --exclude dpdfnet-native
cargo test --locked --profile release-unwind -p dpdfnet-native

The tests use the models in the repository.

pkgbuild/PKGBUILD builds the three packages the same way.

Layout

ops/              int8/SIMD kernels and resampler shared by all engines
deepfilternet3/   DeepFilterNet3 and -LL engines, their shared LADSPA plugin crate,
                  CLIs and benchmark, weight quantizer
silero-vad/       Silero VAD and the voice gate of the denoiser plugins
firered-vad/      FireRedVAD: the AED behind dpdfnet-enhance --speech-only and the
                  streaming model the voice gate can use instead of Silero
limiter-ladspa/   peak limiter
dpdfnet/          DPDFNet engine, model bundle, export tools
gtcrn-aec/        GTCRN-AEC engine, PipeWire SPA plugin, model, evaluation harness
pkgbuild/         Arch/BigLinux split package
testdata/         synthetic speech and the allocation counter shared by the tests
docs/             design notes per engine and performance method

Documentation

• DeepFilterNet3
• DPDFNet
• GTCRN-AEC
• Performance and numerical rules

License

Code: MIT OR Apache-2.0 (LICENSE-MIT, LICENSE-APACHE); the gtcrn-aec crates are Apache-2.0. Model weights keep their upstream licenses. Sources, checksums and changes are listed in NOTICE.