audiosronnx-dpdfnet
ONNX weights for the dpdfnet denoise engine in
audiosronnx — a mirror of
Ceva-IP/DPDFNet (Apache-2.0), pinned so the
library can resolve a fixed revision.
DPDFNet (Dual-Path RNN-based DeepFilterNet) is a streaming speech denoiser built on DeepFilterNet2. Each graph is stateful and consumes one STFT frame at a time:
spec[1, 1, F, 2], state[S] -> spec_e[1, 1, F, 2], state[S]
The Vorbis-windowed STFT/ISTFT runs in numpy inside audiosronnx, so inference is
onnxruntime-only — no torch, and no libdf (unlike the DeepFilterNet3 engine).
| File | Rate | Size |
|---|---|---|
baseline.onnx | 16 kHz | 8.7 MB |
dpdfnet2.onnx | 16 kHz | 10.2 MB |
dpdfnet4.onnx | 16 kHz | 11.6 MB |
dpdfnet8.onnx | 16 kHz | 14.5 MB |
dpdfnet2_8khz.onnx | 8 kHz | 10.2 MB |
dpdfnet8_8khz.onnx | 8 kHz | 14.6 MB |
dpdfnet2_48khz_hr.onnx | 48 kHz | 10.5 MB |
dpdfnet8_48khz_hr.onnx | 48 kHz | 14.9 MB |
The adapter reproduces the upstream reference implementation to max abs err 1.7e-8 (correlation 1.000000) on identical input.
Usage
from audiosronnx import load_denoise
dn = load_denoise("dpdfnet") # dpdfnet2_48khz by default
clean, rate = dn.denoise("noisy_call.wav")
dn = load_denoise("dpdfnet", model="dpdfnet8") # 16 kHz, highest quality
dn = load_denoise("dpdfnet", attn_limit_db=12) # keep a natural noise floor
License
Apache-2.0, inherited from the upstream DPDFNet release.