by amaas

Recurrent neural network training for noise reduction in robust automatic speech recognition

130 Stars 89 Forks Last release: Not found 4 Commits 0 Releases

Available items

No Items, yet!

The developer of this repository has not created any items for sale yet. Need a bug fixed? Help with integration? A different license? Create a request here:


Recurrent neural network training for noise reduction in robust automatic speech recognition.


The software depends on Mark Schmidt's minFunc package for convex optimization, available here: http://www.di.ens.fr/~mschmidt/Software/minFunc.html

Additionally, we have included Mark Hasegawa-Johnson's HTK write and read functions that are used to handle the MFCC files.

We used the aurora2 dataset available here: http://aurora.hsnr.de/aurora-2.html

Getting Started

A sample experiment is in trainauroralocal.m. You must change the first three paths at the top of the file before you can run it. * codeDir: This directory. Where the drdae code is * minFuncDir: Path to the minFunc dependency * baseDir: Where you want to run the experiment. As the experiment runs, intermediate models will be saved in a directory. For simplicity, we found it useful to create separate directories for each experiment There are a number of additional parameters to tune. A few important ones are: * dropout: Enable dropout * tieWeights: Enable tied weights in the network * layerSizes: The sizes of hidden layers in the network and the output layer * temporalLayer: Enables temporal connections in the RNN

Once you have all the parameters tuned, run 'matlab -r trainauroralocal.m'

Using Your Own Datasets

The code is written so that you can try out different datasets by just supplying a different loader. For an example, see load_aurora.m.

We use cookies. If you continue to browse the site, you agree to the use of cookies. For more information on our use of cookies please see our Privacy Policy.