asvspoof2017 icon indicating copy to clipboard operation
asvspoof2017 copied to clipboard

Published 20 hours ago • verified publisher icon azraelkuan
→

Metadata

an implement of asvspoof 2017 using pytorch

Auto Speech Tech Project2

Baseline

  1. mkdir features in the baseline dir
  2. change the data dir pathToDatabase to yourself in baseline_cqcc.m and baseline_mfcc.m
  3. run baseline_cqcc.m like /matlab_dir/matlab -nodisplay -nodesktop -nosplash -r baseline_cqcc, then do same with the baseline_mfcc.m
  4. u will get cqcc and mfcc features in dir features

NNET

in the nnet dir, we use some deep learning algorithm to solve this problem

Result

the column 3 and 4 only use train data the column 5 use train and dev data to test the eval

BaseLine

system feature EER(Dev) EER(Eval) EER(Eval) Frequency Range B Remarks
GMM cqcc 10.35 30.60 24.77 16-8000 96 Baseline !!!

GMM Approach

only one feature use in the gmm

system feature EER(Dev) EER(Eval) EER(Eval) Frequency Range B iter(default 100)
GMM mfcc 14.14 33.08 16-8000 256
GMM mfcc 36.03 36.17 16-2000 256
GMM mfcc 38.60 37.32 2000-4000 256
GMM mfcc 6.86 27.60 4000-8000 256
GMM mfcc 3.53 25.55 6000-8000 256
GMM cqcc 13.44 28.50 16-8000 256
GMM cqcc 40.45 37.98 16-2000 256
GMM cqcc 42.04 39.59 2000-4000 256
GMM cqcc 7.61 27.49 4000-8000 256
GMM cqcc 4.82 20.30 6000-8000 256
GMM cqcc 7.15 19.99 7000-8000 256
GMM cqcc 4.99 18.05 6000-8000 512
GMM cqcc 6.63 18.58 6000-8000 1024
GMM cqcc 5.06 18.32 6000-8000 512 200
GMM cqcc 6.64 18.58 6000-8000 1024 200
GMM cqcc 7.56 18.07 7000-8000 512
GMM cqcc 7.97 17.24 17.64 7000-8000 1024
GMM cqcc 8.11 17.35 17.48 7000-8000 1024 200
GMM cqcc 8.11 17.35 17.48 7000-8000 1024 300

combine all features in the gmm

NNET

system feature EER(Dev) EER(Eval) EER(Eval) Frequency Range B Remarks
LCNN cqcc 11.827 23.919 20.926
LCNN fft 10.298 23.418 17.983
LCNN fft 9.921 22.235 19.365

← Metadata

21
Stars
14
Forks
Watchers

Owner

verified publisher icon azraelkuan

Metadata

an implement of asvspoof 2017 using pytorch

Back