audiosronnx-deepfilternet
DeepFilterNet3 speech denoiser — the three ONNX graphs its own Rust/tract runtime uses,
mirrored for the audiosronnx deepfilternet engine (load_denoise("deepfilternet")).
| graph | inputs | outputs |
|---|---|---|
enc.onnx |
feat_erb[1,1,T,32], feat_spec[1,2,T,96] |
e0..e3, emb, c0, lsnr |
erb_dec.onnx |
emb, e3, e2, e1, e0 |
m[1,1,T,32] ERB mask |
df_dec.onnx |
emb, c0 |
coefs[1,T,96,10] deep-filter taps (freq-major) |
48 kHz. The STFT/ERB frontend and synthesis come from DeepFilterLib (libdf, Rust — no
torch); the mask application and deep filtering run in numpy. Inference is onnxruntime only.
Upstream (unmodified): https://github.com/Rikorose/DeepFilterNet — models/DeepFilterNet3_onnx.tar.gz
Validated: real noisy speech SI-SDR 2.82 -> 11.97 dB (+9.15), corr(clean, denoised) 0.97. Note: denoising degrades ASR (WER 0.533 -> 0.583) — use for perceptual cleanup, not before a recognizer. See the engine's conversion notes.