DeepFilterNet repo skill

Use this skill when a task involves DeepFilterNet, DeepFilterNet2/3, deep filtering for speech/audio enhancement, the Python deepFilter CLI, df.enhance APIs, HDF5 training data, ONNX export/evaluation, or the Rust deep-filter/LADSPA realtime runtime.

First route by task

• Python enhancement and libdf primitives: use python-enhancement for deepFilter, deep-filter-py, df.enhance.init_df, df.enhance.enhance, df.io, model directories, pretrained names, CPU/CUDA selection, and libdf.DF STFT/ERB smoke checks.
• Training data and training setup: use training-data for HDF5 speech/noise/RIR datasets, dataset.cfg, prepare_data, df.train, base_dir, checkpoints/resume, and host-specific batch sizes.
• Model export and evaluation: use model-export-evaluation for ONNX export outputs, model summaries, VoiceBank/DNS2020 evaluation, DNSMOS, PESQ/STOI/composite metrics, and completed export directory validation.
• Rust realtime deployment: use rust-realtime-deployment for the native Rust deep-filter binary, model archives, LADSPA plugin builds, PipeWire virtual source/sink templates, tract/WASM notes, and demo/UI prerequisites.

Confirm install and package surface

Read installation-and-package-map before choosing extras or backend packages. For a safe no-network check from this skill directory:

python scripts/check_deepfilternet_install.py

The checker verifies required imports, deepFilter/deep-filter-py --help, and a tiny libdf STFT/ERB smoke test without downloading models, running training, building Rust, or touching audio services.

DeepFilterNet essentials

• Python package import root: df.
• Main Python CLI entry points: deepFilter and deep-filter-py.
• Required low-level Python extension for enhancement: libdf from deepfilterlib.
• Optional training dataloader extension: libdfdata from deepfilterdataloader.
• Pretrained Python model names used by the package: DeepFilterNet, DeepFilterNet2, DeepFilterNet3; the source default is DeepFilterNet3.
• Python model audio is 48 kHz by default; the package can resample inputs but future agents should make sample-rate decisions explicit.
• CUDA is optional acceleration. CPU is enough for import/API/helper correctness checks unless the user specifically asks to verify GPU performance.
• Rust/LADSPA/PipeWire workflows are system-level: never install Rust packages, write /etc, restart PipeWire, or attach live audio devices without explicit user approval.

Troubleshooting entry point

For cross-cutting failures, start with references/troubleshooting.md. Then move to the nearest sub-skill troubleshooting file for concrete recovery commands and stop conditions.

Provenance and refresh

Read repo-provenance before deciding whether this skill matches a current checkout. If the commit, package metadata, dirty state, source layout, or public entry points differ materially, refresh the repo skill.

Structured router metadata for managed repo-skill import is in repo-routing-metadata.json. This production run intentionally leaves the skill as a self-contained repository artifact and does not import it into a live router.