Skip to content

Latest commit

 

History

History
56 lines (46 loc) · 2.25 KB

File metadata and controls

56 lines (46 loc) · 2.25 KB

Model Training and Evaluation


Main adjustments

  • Swapped GRU blocks + DPGRNN for a psuedo "Grouped-TCN"
  • Removed optional SFE and TRA (mainly for TFLite/LiteRT conversion)
  • Dropped channel and group amounts for quantization compliance
  • Adjusted dilation and padding for quantization compliance

Model directory


Offline Model

Streaming Model

  • wip...

Evaluation Results

  • Dataset: Synthetic DNS3 Test Set
  • Model Params: 19.01 k
  • MMACs: 45.92
  • Examples: See examples
Model Quantization SDR SI-SNR PESQ STOI
Noisy Baseline 3.39 3.39 1.40 0.79
GTCRN-Micro - 10.41 9.85 1.98 0.85
GTCRN-Micro Streaming - ... ... ... ...
GTCRN-Micro Streaming (ONNX) - ... ... ... ...
GTCRN-Micro Streaming (.tflite) Dynamic ... ... ... ...

  • Dataset: DNS3 Blind Test Set
  • Model Params: 19.01 k
  • MMACs: 45.92
Model Quantization DNSMOS-P.808 BAK SIG OVRL
Noisy Baseline 2.96 2.65 3.20 2.33
RNNoise - 3.15 3.45 3.00 2.53
GTCRN (Original) - 3.44 3.90 3.00 2.70
GTCRN-Micro - 3.25 3.60 2.99 2.58
GTCRN-Micro Streaming - ... ... ... ...
GTCRN-Micro Streaming (ONNX) - ... ... ... ...
GTCRN-Micro Streaming (.tflite) Dynamic ... ... ... ...

Acknowledgements

The original model this is based off of is GTCRN, as well as a notable amount of the setup to train and change the model was based off of the same authors project SEtrain.

If you found any of this helpful, please give the original authors repos a star or check out their great work! Without it, none of this project would be possible.