birdyfun / phase-aware-deep-complex-unet

(NOT Official) Implementation DC-UNet (ICLR 2019)

Introuduction

Impelmentation Phase-aware Speech Enhnacement Deep Complex UNet
This is convolution neural networks model for Speech Enhancement
Papers URL

1. Phase-aware Speech Enhancement Deep Complex UNet - openreview
2. Phase-aware Speech Enhancement Deep Complex UNet - arxiv

Architecture

Issue!

Don't use DCUnet16 and DCUnet20 via ComplexBatchNormalization
You can use only Naive_DCUnet series
컴플렉스 배치 정규화를 적용한 DCUnet16, DCUnet20을 사용하지 마세요.
(텐서플로우 최신 버전에서는 위 레이어를 위한 기능이 삭제되었습니다.)
Naive하게 배치 정규화를 사용한 모델만 참고하세요.

Requirements

Python >= 3.6.9 (3.7.6)
numpy
scipy
librosa == 0.7.2 (with numba 0.48.0)
Tensorflow == 2.1.0

Folder

Directory
./Deep_Complex_UNet
    /datasets
        /train_noisy
        /train_clean
        /test_noisy
        /test_clean
    /model_save
    /model_pred
    /complex_layers
        __init__.py
        stft.py
        layer.py
        activation.py
        normaliztion.py
        dcunet.py
    model_module.py
    model_loss.py
    model_data.py
    model_test.py
    model_train.py

Usage

--model naive_dcunet16 naive_dcunet20 dcunet16 dcunet20
--loss SDR or wSDR  
--save "save file name"  
python model_train.py --model dcunet20 --loss wSDR --save dcunet20_default --optim adam --lr 0.002
python model_train.py --model dcunet16 --loss wSDR --save dcunet16_default --optim adam --lr 0.002