Models for our spoofing journal paper: Deep Features for Automatic Spoofing Detection
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.
DNN
RNN
classifier
decoder
README.md
fbank.cfg

README.md

Spoofing_Journal_Models

Models for our recently submitted spoofing journal paper on Speech Communication:

Deep Features for Automatic Spoofing Detection

Yanmin Qian, Nanxin Chen, Kai Yu

We included two best models: DNN model and RNN model.

Dependency:

Steps:

  • You should first extract features using HCopy and fbank.cfg
  • After that you can try BLSTM model and dnn model. Use gzip -d to decompress it.
  • Decoding command: python decode_*.py <decompressed model_file> <scp_file>
  • Classify command: python <classifier.py> <training_set> <cv_set> <development_set> <test_set>

Format for scp file:

name=file_path([start,end])

[start,end] is optional

Format for feature file using for different classifiers(at least required for training set):

label x_1 x_2 ... x_n

Results

Model Classifier S1 S2 S3 S4 S5 S6 S7 S8 S9 S10 known unknown all
DNN LDA 0.03% 0.08% 0.00% 0.00% 0.16% 0.18% 0.02% 0.01% 0.04% 25.47% 0.05% 5.14% 2.60%
DNN GDF 9.50% 28.19% 4.57% 4.36% 36.59% 32.66% 32.44% 7.62% 30.80% 39.97% 16.64% 28.70% 22.67%
DNN SVM 0.20% 0.19% 0.01% 0.03% 0.67% 0.65% 0.09% 0.01% 0.15% 37.12% 0.22% 7.60% 3.91%
DNN OneClassSVM 0.85% 1.03% 0.03% 0.03% 3.69% 3.27% 0.74% 0.01% 0.55% 48.79% 1.12% 10.67% 5.90%
BLSTM LDA 0.01% 0.24% 0.00% 0.00% 0.12% 0.28% 0.17% 0.03% 0.18% 15.28% 0.07% 3.19% 1.63%
BLSTM GDF 0.01% 0.66% 0.00% 0.00% 0.15% 0.51% 0.58% 0.09% 0.52% 19.56% 0.16% 4.25% 2.21%
BLSTM SVM 0.01% 0.85% 0.00% 0.00% 0.26% 0.80% 0.46% 0.03% 0.66% 10.72% 0.22% 2.54% 1.38%
BLSTM OneClassSVM 0.01% 0.68% 0.00% 0.00% 0.18% 0.59% 0.47% 0.07% 0.50% 11.53% 0.17% 2.63% 1.40%

Currently it is under construction.