AOSE_Unet(Audio Only Speech Enhancement using Unet)

0. Prepare dataset 관련(본인 취향의 Dataset을 만들거면 필자처럼 안해도 됨)

1) download

1. download and extract the file train / test dataset 28spk https://datashare.is.ed.ac.uk/handle/10283/2791
2. write address location of files in txt file /home/nas/clean/train.txt /home/nas/noise/train.txt /home/nas/clean/test.txt /home/nas/noise/test.txt

2) STFT the audio data and store it as a pickle.

1. 48khz 데이터들을 16khz로 resampling해준다. 필자는 16khz로 사용했다.
2. python data_feature/demand_data_get_feature.py --clean_train /home/nas/clean/train.txt --noise_train /home/nas/noise/train.txt --clean_test /home/nas/clean/test.txt --noise_test /home/nas/noise/test.txt --train_save_path /home/nas/train/save/ --test_save_path /home/nas/test/save --fs 16
3. or sh make_data_feature.sh

1. train 관련

1. python train_DCUnet_jsdr_demand.py --train_data_root_folder /home/nas/train/save/ --val_data_root_folder /home/nas/test/save --gpu 0 --modelsave_path model/save/path --snr 0 --exp_day 0101 --batch_size 20 --frame_num 128 --learning_rate 0.0001 --fs 48
2. or sh train.sh

2. model 관련

1. frame_num is time value of STFT.
2. recommend frame_num = 128 or 256 like 2's power
3. if you don't like it, change model padding, stride, kernel etc.

3. test 관련

1. python test_DCUnet_jsdr_demand.py --fs 48 --test_model model/path.pth --test_data_root_folder /home/nas/test/audio --test_data_output_path /home/nas/test/output
2. or sh test.sh
3. test_data_root_folder is folder that has .wav audio files.
4. The length of the output audio is limited. To solve this, go intodataset/demand_dataset_test_librosa.py and add line 107-113 paragraphs appropriately to the length.

4. data 추가 관련

1. demand noise dataset download https://zenodo.org/record/1227121#.X1Ytv3kzaUk
2. To do, clean data + demand noise dataset
3. cd data_augment python data_aug_demand_dataset.py --clean_train_txt /home/nas/clean/train.txt --noise_txt /home/nas/demand/noise.txt --save_path /home/nas/save/path/ --snr 0 --fs 48