deepfilternet3_export

DeepFilterNet3 → streaming ONNX export

Exports DeepFilterNet3 as three split ONNX models with explicit GRU hidden-state I/O, enabling chunk-by-chunk inference in C#.

The standard single-call export processes all T frames at once, which gives poor GPU utilisation because GRU layers must step sequentially. The streaming export processes audio in chunks of ~100 frames (~1 s), carrying GRU state across chunks.

Files

• export_df3_streaming.py — exports the three streaming ONNX models and writes streaming_meta.json
• requirements.txt — export dependencies

Environment

pip install -r scripts/deepfilternet3_export/requirements.txt

Export

python scripts/deepfilternet3_export/export_df3_streaming.py \
  --out external/deepfilternet_onnx_streaming \
  --opset 14

Outputs:

• enc.onnx
• erb_dec.onnx
• df_dec.onnx
• streaming_meta.json — hidden-state shapes and chunk size

ONNX Contract

enc.onnx

Name	Shape	Description
feat_erb	[1, 1, T, 32]	ERB filterbank features
feat_spec	[1, 2, T, 96]	Complex spectrogram features
h_enc	[1, 1, 256]	Encoder GRU hidden state in
e0..e3	[1, 64, T, *]	Skip connections
emb	[1, T, 512]	Embedding
c0	[1, 64, T, 96]	DF conv features
lsnr	[1, T, 1]	Log SNR estimate
h_enc_out	[1, 1, 256]	Encoder GRU hidden state out

erb_dec.onnx

Name	Shape	Description
emb	[1, T, 512]	Embedding (from enc)
e3..e0	[1, 64, T, *]	Skip connections (from enc)
h_erb	[2, 1, 256]	ERB decoder GRU hidden state in
m	[1, 1, T, 32]	ERB mask
h_erb_out	[2, 1, 256]	ERB decoder GRU hidden state out

df_dec.onnx

Name	Shape	Description
emb	[1, T, 512]	Embedding (from enc)
c0	[1, 64, T, 96]	DF conv features (from enc)
h_df	[2, 1, 256]	DF decoder GRU hidden state in
coefs	[1, T, 96, 10]	Deep filtering coefficients
h_df_out	[2, 1, 256]	DF decoder GRU hidden state out

T is a dynamic axis — use any chunk size consistent with streaming_meta.json.