This is the repository of the RSRGAN project. Our original paper can be found here.
In this work we investigate the use of generative adversarial networks (GANs) in speech dereverberation for robust speech recognition.
Our RIRs were from here, and we used KALDI to simulate reverberant speech.
All the project was developed with TensorFlow and KALDI and our project was based on SEGAN.
- Python 2.7
- TensorFlow 1.4.0
- KALDI
-
You should prepare your data according KALDI format as following:
data |- wav.scp |- ...
then, runing the following command to simulate reverbrant data.
# remember to change the data path in run.sh bash reverberate/run.sh
-
After this step, you can get reverbrant waves and their counterparts, i.e., the clean speeches. Using KALDI command
compute-spectrogram-feats
to get log-power spectrum (LPS) features as your inputs. Usingcompute-mfcc-feats
to get MFCC feature as your labels. In our experiments, we used 257-dim LPS and 40-dim MFCC without delta and delta-delta.- LPS: hamming window
- MFCC: default config file in KALDI WSJ recipe (conf/mfcc_hires.conf)
-
Training GAN
# shell file name with 'placeholder' is the script that training 'G' and 'D' using the same min-batch's data. # remeber to change the data path in shell files # you can also try other shell files to reproduce the other experimental results. (E.g., DNN, RCED, LSTM) bash run_gan_rnn_placeholder.sh
- You should have a well-trained ASR model and make sure your AM's input feature is the same with GAN generator's output feature.
- Then you can use any kind of decoder to decode enhanced feature.
If the code of this repository was useful for your research, please cite our work:
@article{ke2018gan,
title = {Investigating Generative Adversarial Networks based Speech Dereverberation for Robust Speech Recognition},
author = {Ke, Wang and Junbo, Zhang and Sining, Sun and Yujun, Wang and Fei Xiang, and Lei, Xie},
journal = {arXiv preprint arXiv:1803.10132},
year = {2018},
}
e-mail: [email protected]