DeepFilterNet3 - KICK ONNX (single file, fixed 1500 frames)
The three official DeepFilterNet3 ONNX networks (enc.onnx, erb_dec.onnx, df_dec.onnx) joined by KICK into one ONNX graph, so that
KICK Studio can run it on Windows and macOS with ONNX Runtime (CPU). The weights are unchanged (fp32, not compressed); only the graph is joined.
Interface (deepfilternet3_t1500.onnx, 8.6 MB)
| name | shape | |
|---|---|---|
| in | feat_erb |
[1, 1, 1500, 32] |
| in | feat_spec |
[1, 2, 1500, 96] |
| in | at |
int64 scalar: frame after which the GRU states are returned |
| in | h0 ... h4 |
[1, 1, 256] GRU states (0 = enc, 1-2 = erb_dec, 3-4 = df_dec) |
| out | m |
[1, 1, 1500, 32] ERB mask |
| out | coefs |
[1, 1500, 96, 10] deep-filter coefficients |
| out | s0 ... s4 |
[1, 1, 256] GRU states after frame at, to carry into the next chunk |
Feature extraction (STFT 960/480, ERB, deep filtering) is done outside the network, as in the original libDF.
Checks
- With zero states the outputs are bit-identical to the three original files.
- Carrying the states across two chunks matches one long run (0.1 % difference against 18 % without carrying).
- MacBook Air M4, ONNX Runtime CPU: 0.07 s for 15 s of mono audio. Synthetic speech + white noise: SNR improves by 6 to 13 dB.
Credits and license
- Model and weights: Hendrik Schröter - Rikorose/DeepFilterNet, DeepFilterNet3
(
models/DeepFilterNet3_onnx.tar.gz). Paper: DeepFilterNet: Perceptually Motivated Real-Time Speech Enhancement. - Dual-licensed MIT / Apache-2.0 like the upstream project (see
LICENSE-MIT,LICENSE-APACHE).