DPCRN_DNS3

Created on Mon Oct 28 16:05:31 2021
@author: xiaohuai.le

This repository is the official implementation of paper "DPCRN: Dual-Path Convolution Recurrent Network for Single Channel Speech Enhancement". This work got the third place in Deep Noise Suppression Challenge.

Requirements

tensorflow>=1.14,
numpy,
matplotlib,
librosa,
sondfile.

Datasets

We use Deep Noise Suppression Dataset and OpenSLR26, OpenSLR28 RIRs dataset in our training and validation stages. The directory structure of the dataset is shown below:
dataset
├── clean
│ ├── audio1.wav
│ ├── audio2.wav
│ ├── audio3.wav
│ ...
├── noise
│ ├── audio1.wav
│ ├── audio2.wav
│ ├── audio3.wav
│ ...

RIR
├── rirs
│ ├── rir1.wav
│ ├── rir2.wav
│ ├── rir3.wav
│ ...

Training and test

Run the following code to training:

python main.py --mode train --cuda 0 --experimentName experiment_1

Run the following code to test the model on a single file:

python main.py --mode test --test_dir the_dir_of_noisy --output_dir the_dir_of_enhancement_results

More samples

The final results on the blind test set of DNS3 is available on https://github.com/Le-Xiaohuai-speech/DPCRN_DNS3_Results.

Real-time inference

Note that the real-time inference can only run on the tensorflow=1.x. Run real-time inference to calculate the time cost of a frame:

python ./real_time_processing/real_time_DPCRN.py

Tensorflow Lite quantization and pruning

The TFLite file of a smaller dpcrn model is uploaded. Enhance a single wav file:

python ./inference/real_time_inference/inference.py

Streaming recording and enhancement:

python ./inference/real_time_inference/recording.py

Citations

@inproceedings{le21b_interspeech,
  author={Xiaohuai Le and Hongsheng Chen and Kai Chen and Jing Lu},
  title={{DPCRN: Dual-Path Convolution Recurrent Network for Single Channel Speech Enhancement}},
  year=2021,
  booktitle={Proc. Interspeech 2021},
  pages={2811--2815},
  doi={10.21437/Interspeech.2021-296}
}