/ulaw-SGAN-for-SE

ulaw SGAN for speech enhancemen

Primary LanguagePython

code of paper Li, H., Xu, Y., Ke, D., & Su, K. (2021). μ-law SGAN for generating spectra with more details in speech enhancement. Neural Networks, 136, 17-27.

Prepare for running

  1. Running bash ./nn_se/_1_perprocess.sh to prepare data.

  2. Change "root_dir" parameter in nn_se/FLAGS.py to the root of the project. For example "root_dir = /home/user/ulaw-SGAN-for-SE".

  3. Ensure "PARAM = dse_ulawV2_G_FTmagmse_Ndloss_ssnr_001" is set in last line of nn_se/FLAGS.py.

  4. Running cp nn_se dse_ulawV2_G_FTmagmse_Ndloss_ssnr_001 -r to create the Experiment config code dir.

Train

Running python -m dse_ulawV2_G_FTmagmse_Ndloss_ssnr_001._2_train to start training of config "dse_ulawV2_G_FTmagmse_Ndloss_ssnr_001".

Evaluate

Running python -m dse_ulawV2_G_FTmagmse_Ndloss_ssnr_001._3_enhance_testsets to get the metrics of Experiment "dse_ulawV2_G_FTmagmse_Ndloss_ssnr_001". The last ckpt is selected as the default ckpt to load. Alse, you can use --ckpt to specify the path of ckpt.

More

See "nn_se/_1_preprocess.sh", "nn_se/_2_train.py" and "nn_se/_3_enhance_testsets.py".