Skip to content

README.md

Latest commit

 

History

History
74 lines (58 loc) · 2.54 KB

README.md

File metadata and controls

74 lines (58 loc) · 2.54 KB
Raw

DeepFilterNet

Model infomation

This folder provides the code needed to reproduce the DeepFilterNet noise suppresion model described in the DeepFilterNet: A Low Complexity Speech Enhancement Framework for Full-Band Audio based on Deep Filtering paper.

We modified the model to cut off output postprocessing and fix input time dimension to 200, then convert ONNX format based on tensorflow-onnx tools.

For evaluation, we use dataset from https://datashare.ed.ac.uk/handle/10283/2791 same as RNNoise example.

Name Task Source FP32 W8A8
DeepFilterNet Noise Suppresion https://github.com/Rikorose/DeepFilterNet Model/sim_deepfilter_200.onnx Model/model.tflite
Precision Mode Format Metric (100 wav files PESQ) NN SDK Version
FP32 ONNX (opset_v13) 2.82 N/A
W8A8 TFLite v2.17.0 2.793 (AndesQuant result) v1.1.1

License

Dataset build

The dataset is accessible for download at the following link:

In the website download following as validation data:

  • clean_testset_wav.zip (147.1Mb)
  • noisy_testset_wav.zip (162.6Mb)

Download following as training data:

  • clean_trainset_56spk_wav.zip (4.442Gb)
  • noisy_trainset_56spk_wav.zip (5.240Gb)

Extract the zip file and the final structure show as following:

/your_path/rnnwave/
    ├── clean_trainset_56spk_wav/
    │   ├── p234_001.wav
    │   ├── p234_002.wav
    │   ├── p234_003.wav
    │
    ├── noisy_trainset_56spk_wav/
    ├── clean_testset_wav/
    └── noisy_testset_wav/

Each of the clean and noisy folders for the train and test data will contain multiple .wav audio files.

Change model_cfg.yaml

Revise path in followin model_cfg.yaml to your path

train_clean_wave_path: "/your_path/rnnwave/clean_trainset_56spk_wav/"
train_noisy_wave_path: "/your_path/rnnwave/noisy_trainset_56spk_wav/"
clean_wave_path: "/your_path/rnnwave/clean_testset_wav"
noisy_wave_path: "/your_path/rnnwave/noisy_testset_wav"
dummy_input: [[1,1,42],[1,24],[1,48],[1,96]]
batch_size: 1
channel: 1
height: 42
width: 1
fp32_min: -41.1027
fp32_max: 41.2633

The dataset setting is done.

Available workflow

Prun SVD PTQ QAT
Symm x v v x
Asym x v v x