Light

wangkenpu / rsrgan Goto Github PK

View Code? Open in Web Editor NEW

58.0 10.0 16.0 3.41 MB

Robust Speech Recognition Using Generative Adversarial Networks (GAN)

License: MIT License

Python 33.74% Shell 56.43% Perl 9.82% sed 0.01%

rsrgan's Introduction

Robust Speech Recognition Using Generative Adversarial Networks (GAN)

Introduction

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.

Dependencies

Python 2.7
TensorFlow 1.4.0
KALDI

Date preparation and GAN training

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. Using compute-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

ASR decoding

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.

Reference

If the code of this repository was useful for your research, please cite our work:

@inproceedings{wang2018investigating,
  author    = {Wang, Ke and Zhang, Junbo and Sun, Sining and Wang, Yujun and Xiang, Fei and Xie, Lei},
  booktitle = {Interspeech 2018},
  pages     = {1581--1585},
  title     = {{Investigating Generative Adversarial Networks based Speech Dereverberation for Robust Speech Recognition}},
  year      = {2018}
}

Contact

e-mail: [email protected]

rsrgan's People

Contributors

Stargazers

Watchers

Forkers

linan2 rpersie willqucd twistedmove entn-at aascode ml-lab fendaq xinkez byfaith okrio lbxcfx vyoz 5iding xuanphu108 mohmdsh

rsrgan's Issues

Where can I get training data?

Where can I get training data?

performance degrade using aishell as training data of feature-mapping

I tried lstm/res-lstm/gan-res-lstm, use the same configurations, all experiments got performance degrading. I don't know what's wrong. The back-end asr system is tdnn+lstm. front-end is feature mapping from aishell_train_clean+rvb to aishell_train_clean . Do you have any insights ? thank you very much !

mulit-GPU Performs worse than single-GPU

Dear wang,
First, thanks for your code.
I use your code to do experiments, but I find the mulit-GPU Performs worse than single-GPU. Do you suffer from this problem? Or can you give me some advices ? I just have to change the
batch size and learning rate ?

Thank you !

preparing dataset

I was wonderful to know that this repository gives considerable speech reverberation result.
So I've tried to train reverb model with LibiSpeech corpus.
I've all done making reverberation wave dataset.
But I dont well know how I can make mfcc and lps dataset.
I've used the script in calculation of mfcc like this:
"""
if [ $stage -le 6 ]; then
for part in dev_clean test_clean dev_other test_other train_clean_100; do
steps/make_mfcc.sh --cmd "$train_cmd" --nj 40 data/$part exp/make_mfcc/$part $mfccdir
steps/compute_cmvn_stats.sh data/$part exp/make_mfcc/$part $mfccdir
done
fi
"""

and i've replaced "compute-mfcc-feats" as "compute-spectrogram-feats" for LPS.
Then I am trying to use cmvn_train.ark cmvn_train_reverb.ark of "mfcc " directory as labels.cmvn and inputs.cmvn.
but I've got following error on stage 0.
"""
Prepare tr and cv data
Make Numpy format Global CMVN file ...
Convert data/train/inputs.cmvn and data/train/labels.cmvn to Numpy format
Input .ark file is not binary
"""

I dont well know also how I can make cv.list, train.list.
It would be greatly appreciated if you could tell me how to make it.
Thanks
Ting

performance degrade

I tried lstm/res-lstm/gan-res-lstm, use the same configurations, all experiments got performance degrading. I don't know what's wrong. The back-end asr system is tdnn+lstm. front-end is feature mapping from aishell_train_clean+rvb to aishell_train_clean . Do you have any suggestions or insights ? thank you very much !

test

Some questions about datatset？

Hi,
I have read your paper about this repository. I did learn a lot. Now, I want to employ your code to make some experiments. I have some questions and hope you can give me some advice.
(1) Do you use GAN in the condition( both reverberant and noisy)？ What's the results？
(2) Do I need to prepare a clean data before simulating the reverb data? However ， My data is collected form telephone and mobile phone.

Recommend Projects

React

A declarative, efficient, and flexible JavaScript library for building user interfaces.
Vue.js

🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
Typescript

TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
TensorFlow

An Open Source Machine Learning Framework for Everyone
Django

The Web framework for perfectionists with deadlines.
Laravel

A PHP framework for web artisans
D3

Bring data to life with SVG, Canvas and HTML. 📊📈🎉

Recommend Topics

javascript

JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
web

Some thing interesting about web. New door for the world.
server

A server is a program made to process requests and deliver data to clients.
Machine learning

Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
Visualization

Some thing interesting about visualization, use data art
Game

Some thing interesting about game, make everyone happy.

Recommend Org

Facebook

We are working to build community through open source technology. NB: members must have two-factor auth.
Microsoft

Open source projects and samples from Microsoft.
Google

Google ❤️ Open Source for everyone.
Alibaba

Alibaba Open Source for everyone
D3

Data-Driven Documents codes.
Tencent

China tencent open source team.