Feature extraction using unsupervised data-driven modulation filtering approach
Switch branches/tags
Nothing to show
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Failed to load latest commit information.
Five_example_feature_extraction_speech_forASR.m
Four_hidden_prob_avg.m
LICENSE
One_example_filtLearning_Rate_crbm.m
README
README.txt
Three_example_filtSelection_validation.m
Two_example_filtLearning_Scale_crbm.m
Zero_example_mel_spectrogram_speech.m
aud2cor_dataMod_rate.m
aud2cor_dataMod_scale.m
conve.m
convemex.mexa64
convs.m
convs4.m
convs4mex.mexa64
convsmex.mexa64
crbm_filtSelection_validation.m
data.tar.gz
getparams.m
getparams_rate.m
getparams_scale.m
make.m
make_input_matrix.m
make_input_matrix_Scale.m
make_input_matrix_dev.m
make_input_matrix_one_utt.m
melSpec.tar.gz
melSpec_codes.tar.gz
mex.tar.gz
modFilt_spec.tar.gz
nvmex.m
nvmex_helper.m
poolHidden.m
poolHmex.mexa64
sourceFile_wav_finalFeat.txt
sourceFile_wav_spec.txt
trainCRBM_Rate.m
trainCRBM_Scale.m
validation_spec.tar.gz
validation_spec.txt
wav_file.tar.gz

README

%******************************************************************
% Purvi Agrawal and Sriram Ganapathy
% Learning and Extraction of Acoustic Patterns (LEAP) Lab
% Indian Institute of Science, Bangalore, India
% {purvi_agrawal,sriram}@ee.iisc.ernet.in
%******************************************************************
% 31-Aug-2017
% See the file LICENSE for the licence associated with this software.
%******************************************************************
Ref:
1. P. Agrawal, S. Ganapathy, “Unsupervised Modulation Filter Learning for Noise-Robust Speech Recognition,” Journal of Acoustical Society of America, EL, 2017.
******************************************************************
Description:

The folder contains the MATLAB codes to extract features using unsupervised data-driven modulation filtering approach.

- The modulation filters (rate and scale separately) are learned from mel spectrograms 
  using Convolutional Restricted Boltzmann Machine (CRBM).
- Multiple filters are learned using residual approach.
- The filter selection criteria uses average hidden activation probability values.
- The mel spectrograms are filtered using the selected modulation filters 
  which are used as features.

Sequence to run :

0. Zero_example_mel_spectrogram_speech.m
1. One_example_filtLearning_Rate_crbm.m
2. Two_example_filtLearning_Scale_crbm.m
3. Three_example_filtSelection_validation.m
4. Four_hidden_prob_avg.m
5. Five_example_feature_extraction_speech_forASR.m

The reference wav files, spectrograms and few codes are saved in compressed folders.

Acknowledgements:

%   Some of the functions in this implementation modify the original implementation by PENG QI
%   Original code can be found at : http://qipeng.me/software/convolutional-rbm.html
%   https://github.com/qipeng/convolutionalRBM.m. 
%   For research / personal purposes only.