@Shitao Weng: sweng@andrew.cmu.edu
@Shushan Chen: shushanc@andrew.cmu.edu
CUDA 6.5
CMake 2.8+
g++ (c++11 stardard)
git clone .
cd cuda-mfcc-gmm/
mkdir build/
cmake ../
make && make install
cd cuda-mfcc-gmm/RunEnv
#./cpu_mfcc -l wav_file_name_without_prefix
./cpu_mfcc -l SW_20001_ch2_cut
Stardard Output Stream:
PreEmp: 1.0210488 s , 1.36219%
Windowing: 0.0832372 s , 5.38676%
FFT padding: 0.018985 s , 1.22863%
PowerSpectrum: 0.586775 s , 37.9736%
MelFiltering: 0.434919 s , 28.1461%
DCT Ceptrum: 0.39311 s , 25.4404%
Normalization: 0.00714397 s , 0.462328%
MFCC feature will be stored in file normalMelCeps.txt.
#./cuda_mfcc -l wav_file_name_without_prefix
./cuda_mfcc -l SW_20001_ch2_cut
Stardard Output Stream:
CUDA Initialize Time: 2.32246
Total Time (Without InitializeTime) : 0.070796
PreProcessing: 0.00680399 s , 9.6107%
FFT: 0.051362 s , 72.5493%
MelFiltering: 0.00560689 s , 7.91978%
DCT Ceptrum: 0.00620914 s , 8.77046%
Normalization: 0.000813961 s , 1.14973%
MFCC feature will be stored in file cuda_normalMelCeps.txt.
Comparing the running of each stages used by CPU and GPU, we get a 25.5x speedup without considering the CUDA initialization time.
diff normalMelCeps.txt cuda_normalMelCeps.txt
# ./gmm_train_main config_file ouput_model_file_name
./gmm_train_main ubm.64.cfg ubm.cpu.64.model
*************Split MixNum = 32 Finished ******************
Avg EM Iteration: 0.768752
Avg EM time 2.306273
Avg MixUp time 64, 0.000240
Avg KMean time 0.843861
Avg KMean Iteration: 0.281287
*************Split MixNum = 64 Finished ******************
Avg EM Iteration: 1.553885
Avg EM time 18.646647
all training has been finished
Last Gmm MixtureNum=64
Whole runtime : 25.036303
# ./cuda_gmm_train_main config_file ouput_model_file_name
./cuda_gmm_train_main ubm.64.cfg ubm.cuda.64.model
...
*************Split MixNum = 32 Finished ******************
Avg EM Iteration: 0.039880
Avg EM time 0.119680
Avg MixUp time 64, 0.000319
Avg KMean time 0.004120
Avg KMean Iteration: 0.001373
*************Split MixNum = 64 Finished ******************
Avg EM Iteration: 0.070802
Avg EM time 0.849650
all training has been finished
Last Gmm MixtureNum=64
Whole runtime : 3.408745
On this small test, we get 7x speedup(100 wav files, 21420 frames). If we test it on a big data set (4000 wav files, 971526 frames), we can get 27x speed up.
Before you start running this, please download the mfcc features files from here: mfcc.zip.
cd cuda-mfcc-gmm/RunEnv
unzip mfcc.zip
./cuda_gmm_train_main ubm.512.cfg ubm.cuda.512.model
./svm_feature_extract ubm.cuda.512.model all.list all.tags 16375
This will generate two libsvm feature files svm_train.feature and svm_test.feature.
After generating feature files for libsvm, we can use libsvm to classify the wav files as human utterance or spoofed utterance.
./svm-train -s 0 -t 1 -d 3 -b 1 -g 1 -r 1 svm_train.feature spoof.model
./svm-predict -b 1 svm_test.feature spoof.model result.txt
Accuracy = 85.1083% (45424/53372) (classification)]
The MIT License (MIT)
Copyright (c) <2015> Shitao Weng(sweng@andrew.cmu.edu), Shushan Chen(shushanc@andrew.cmu.edu)
Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.