tested working on python 3.7 environment
pip install numpy soundfile spafe sklearn
The code files are divided into three categories: Code for Single Feature GMM Base Model, Code for Frame-level (ensemble) Fusion GMM model, Code for Instance-level / File-level Fusion GMM model.
The tDCF_python/ is a official tool published by ASVspoof 2021 to calculate Equal Error Rate (EER) and min t-DCF score.
For all the code file below, please checkout specific argument format with -h flag.
gmm_data_processing.py will extract cepstral coefficients from audio file data and store them in Python pickle file under folder data/
gmm_train_recur.py will train GMM base model chunk-by-chunk recursively, since training set will not fir into memory for most computers, the trained GMM model will be pickled and stored in model/
gmm_test.py will test the trained GMM model on evaluation dataset and stored an specific format text in output/ for t-DCF and EER metrics.
preprocess.sh, train.sh, and test.sh are Shell script to help you run data processing, training, and testing on base GMM model.
gmm_ensemble_preprocessing.py will take an specific GMM base model and a dataset, generate the base model's score output, and then store them into pickle files for frame-level fusion use in ensemble/
gmm_ensemble_train.py and gmm_ensemble_train_recur.py will train the frame-level fusion GMM model based on data provided. The normal train will handle smaller size GMM models, while the recur train file is needed when training large number component GMM models.
gmm_ensemble_test.py will test the trained frame-level GMM model on evaluation dataset and stored an specific format text in output/ for t-DCF and EER metrics.
gmm_filelevel_preprocessing.py will take an specific frame-level fusion GMM model and a dataset, generate the frame-level fusion model's score output, and then store them into pickle files for instance-level fusion use in ensemble/. Later, manually add or remove the frame-level model's output score you would like to fuse into a new instance-level model. Generally, the data for instance-level is stored in experiment/.
gmm_filelevel_train.py will train the instance-level fusion GMM model based on data provided and store the model in model/.
gmm_filelevel_test.py will test the trained instance-level GMM model on evaluation dataset and stored an specific format text in output/ for t-DCF and EER metrics.
silence_measure.py will calculate the leading and trailing silence frame count, in order to generate the 'silence' feature. This functions can be invoked when passing --ftype silence into data preprocessing Python file.
score_eval.py is a wrapper program over tDCF_python to help calculate evaluation metrics, specific score text generated by various *_test.py can be read from output/
Assuming ASVspoof 2019's Logical Access dataset is located at ./LA
Take MFCC as an example: assume LA dataset is located at /mnt/LA
echo "extracting features: mfcc"
echo "preparing training data"
python3 gmm_data_processing.py --data_split train \
--data_path ./LA/ASVspoof2019_LA_train/flac \
--output_path ./data/train/ \
--label_path ./LA/ASVspoof2019_LA_cm_protocols/ASVspoof2019.LA.cm.train.trn.txt \
--ftype mfcc
echo "preparing dev data"
python3 gmm_data_processing.py --data_split dev \
--data_path ./LA/ASVspoof2019_LA_dev/flac \
--output_path ./data/dev/ \
--label_path ./LA/ASVspoof2019_LA_cm_protocols/ASVspoof2019.LA.cm.dev.trl.txt \
--ftype mfcc
echo "preparing test data"
python3 gmm_data_processing.py --data_split eval \
--data_path ./LA/ASVspoof2019_LA_eval/flac \
--output_path ./data/test/ \
--label_path ./LA/ASVspoof2019_LA_cm_protocols/ASVspoof2019.LA.cm.eval.trl.txt \
--ftype mfcc
The saved pickle file has the format: [(feature[framecountx60], label[bonafide/spoof]) x N instances]
PLEASE save data pickles under directories named with feature type.
After preprocessing to pickles, please scp pickles to /mnt/LA/preprocessed/...
Take MFCC as an example:
echo "training on features: mfcc"
python3 gmm_train.py --data_path ./data/train/mfcc.pkl \
--model_path ./model/mfcc
PLEASE save model pickles under directories named with feature type. After training to pickles, please scp pickles to /mnt/LA/models/...
Take MFCC as an example:
echo "testing on features: mfcc"
python gmm_test.py --data_path ./data/eval/eval-mfcc.pkl \
--model_path_bon ./model/mfcc/bon.gmm \
--model_path_sp ./model/mfcc/sp.gmm \
--label_path ./LA/ASVspoof2019_LA_cm_protocols/ASVspoof2019.LA.cm.eval.trl.txt \
--output_path ./output/mfcc-eval.text
python score_eval.py --cm_score ./output/mfcc-eval.text \
--asv_score ./tDCF_python/scores/ASVspoof2019_LA_eval_asv_scores.txt
Take LFCC as an example:
python gmm_ensemble_preprocess.py --data_split eval --data_path ./data/eval/eval-lfcc.pkl --model_path ./model/lfcc --model_name lfcc --output_path ./ensemble/lfcc
Take LFCC as an example:
python gmm_ensemble_train_recur.py --data_dir ./experiment/lfcc/train --model_path ./model/ensemble-lfcc
Take LFCC as an example:
python gmm_filelevel_preprocess.py --data_split train --data_dir ./experiment/lfcc/train --model_path ./model/ensemble-lfcc --output_path ./ensemble/ensemble-lfcc --model_name ensemble-lfcc
Take LFCC as an example:
python gmm_ensemble_preprocess.py --data_split eval --data_path ./data/eval/eval-lfcc.pkl --model_path ./model/lfcc --model_name lfcc --output_path ./ensemble/lfcc
Take LFCC as an example:
python gmm_filelevel_train.py --data_dir ./experiment/filelevel-lfcc/train --model_path ./model/filelevel-lfcc
Take LFCC as an example:
python gmm_filelevel_test.py --data_dir ./experiment/filelevel-lfcc/eval --model_path_bon ./model/filelevel-lfcc/bon.gmm --model_path_sp ./model/filelevel-lfcc/sp.gmm --label_path ./LA/ASVspoof2019_LA_cm_protocols/ASVspoof2019.LA.cm.eval.trl.txt --output_path ./output/filelevel-lfcc.txt