YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
DeepFilterNet 3 ONNX
Mirror of the DeepFilterNet 3 ONNX models from
github.com/Rikorose/DeepFilterNet,
specifically DeepFilterNet3_onnx.tar.gz extracted.
License
Dual MIT / Apache-2.0 — see upstream repository.
Files
| File | Role | Size |
|---|---|---|
enc.onnx |
Encoder | 1.9 MB |
erb_dec.onnx |
ERB band decoder | 3.3 MB |
df_dec.onnx |
Deep filtering decoder | 3.3 MB |
config.ini |
Model + runtime configuration | 2 KB |
Total ~8.5 MB on disk.
Runtime contract
DeepFilterNet 3 is a streaming, real-time speech enhancement model:
- Native sample rate: 48 kHz (the model can also handle 16 kHz via internal resampling done by the runtime)
- Frame: 480 samples @ 48 kHz (10 ms)
- Latency: ~40 ms algorithmic
- 3-stage pipeline: STFT → encoder → ERB band gains + deep filter coefficients → iSTFT
Best loaded through the upstream reference implementation (Rust or Python) which handles the STFT, framing, state management, and dual-decoder fusion. The bare ONNX files do not include those.
Why this mirror exists
iOS SDK demo pipeline needs a stable, fast download URL pinned by SHA-256.
The upstream models/ folder URL changes between commits; this mirror gives
us reproducibility for the SDK's SherpaOnnxSpeechEnhancerConfiguration.