Deep Complex U-Net

---

Unofficial PyTorch Implementation of Phase-Aware Speech Enhancement with Deep Complex U-Net, (H. Choi et al., 2018)

Note

This is NOT author's implementation.

Architecture

---

(TO BE)

Requirements

torch==1.1
soundfile==0.9.0
easydict==1.9
git+https://github.com/keunwoochoi/torchaudio-contrib@61fc6a804c941dec3cf8a06478704d19fc5e415a
git+https://github.com/sweetcocoa/PinkBlack@e45a65623c1b511181f7ea697ca841a7b2900f17
torchcontrib==0.0.2
git+https://github.com/vBaiCai/python-pesq
# gputil # if you need to execute multiple training process

Train

---

1. Download Datasets:

• https://datashare.is.ed.ac.uk/handle/10283/2791

2. Separate each train / test wavs

3. Downsample wavs

# prerequisite : ffmpeg.
# sudo apt-get install ffmpeg (Ubuntu)
bash downsample.sh   # all wavs below $PWD will be converted to .flac, 16k samplerate

4. Train

python train_dcunet.py --batch_size 12 \
                       --train_signal /path/to/train/clean/speech/ \
                       --train_noise /path/to/train/noisy/speech/ \
                       --test_signal /path/to/test/clean/speech/ \
                       --test_noise /path/to/test/noisy/speech/ \
                       --ckpt /path/to/save/checkpoint.pth \
                       --num_step 50000 \
                       --validation_interval 500 \
                       --complex

# You can check other arguments from the source code. ( Sorry for the lack description. )                        

Test

---

python estimate_directory.py --input_dir /path/to/noisy/speech/ \
                             --output_dir /path/to/estimate/dir/ \
                             --ckpt /path/to/saved/checkpoint.pth

Results

---

PESQ(cRMCn/cRMRn)	Paper	Mine*
DCUNet - 10	2.72/2.51	3.03/3.07
DCUNet - 20	3.24/2.74	3.12/3.11

• cRMCn : Complex-valued input/Output
• cRMRn : Real-valued input/Output

Comparing the two(Paper's, Mine) values above is inappropriate for the following reasons:

• * I did not use matlab code that the author used to calculate pesq, but instead used pypesq.

• * The Architecture of model is slightly different from the original paper. (Such as kernel size of convolution filters)

• MelSpec

Notes

---

• Log amplitute estimate has slightly worse performance than non-log amplitude
• Complex-valued network does not make the metric better..

Sample Wavs

---

Mixture	Estimated Speech	GT(Clean Speech)
mixture1.wav	Estimated1.wav	GroundTruth1.wav
mixture2.wav	Estimated2.wav	GroundTruth2.wav

Contact

---

• Jongho Choi(sweetcocoa@snu.ac.kr / Seoul National Univ., ESTsoft )