Large Replay Parallel Dataset

This repository contains code for experiments described in paper.

Dependencies

Install dependencies by running pip install -r requirements.txt

Data

Before running train, one should acquire datasets:

• LRPD dataset using link
• ASVSpoof2017 dataset using provided script by running python utils/load_asv17.py OUTPUT_FOLDER or accessing it on official web page
• MUSAN,DECASE 2017 Task3 Acoustic Scenes,DEMAND datasets as noise datasets

And setup pathes to dataset roots in data.yml:

lrpd_root: /PATH/TO/LRPD
asv17_root: /PATH/TO/ASVSpoof2017
noise_roots: 
    - /PATH/TO/MUSAN/ [optional]
    - /PATH/TO/DEMAND/ [optional]
    - /PATH/TO/DECASE/ [optional]
    ...

ADD:

• Description of available model architectures in model_config.json

Training

• Export current path to environment

export PYTHONPATH=$(pwd)

To run training refer to: python3 train.py --help:

usage: train.py [-h] [--train_config TRAIN_CONFIG] [--model_config MODEL_CONFIG] [--gpus GPUS] [--exp_dir EXP_DIR] task dataset_setup

positional arguments:
  task                  antispoofing or device_detector
  dataset_setup         Dataset setup in form 'lrpd_office,lrpd_aparts,asv17_train'

optional arguments:
  -h, --help            show this help message and exit
  --train_config TRAIN_CONFIG
                        Path to train config serialized into JSON (default: /media/ssdraid0cgpu01/home/iiakovlev/new-pipeline/audio-pipelines-pytorch/configs/common/train_config.json)
  --model_config MODEL_CONFIG
                        Path to model config serialized into JSON (default: /media/ssdraid0cgpu01/home/iiakovlev/new-pipeline/audio-pipelines-pytorch/configs/common/model_config.json)
  --gpus GPUS           IDs of GPUs to train on. For example : 0,1 (default: 0)
  --exp_dir EXP_DIR     Path to experiment folder (default: None)

For example: python3 train.py antispoofing lrpd_office,lrpd_aparts,asv17_train will run training of antispoofing detector using LRPD-office, LRPD-aparts and ASVSpoof2017 train part as training data