A torch based open-source framework for detecting audio generated from generative systems. The overarching goal of this repository is to be a toolbox for datasets, augmentations, and pretrained baseline models to assist researchers in the task of audio deepfake detection.
To install, simply use pip:
pip install audio-sleuth
The goal of this package is to contain augmentations, models, and losses all as a torch.nn.Module class. Each augmentation contains an internal LabelAlignment such that you can implement various time-frequency transformations to automatically translate the sample-wise labels as well.
Here is an example of how to chain together augmentations on the Half Truth Dataset.
from audio_sleuth.augmentations import Resample, LFCC, Augmentations
from audio_sleuth.data.datasets import pad_and_transform_collate, HalfTruthDataset
from torch.utils.data import DataLoader
# Create chain of transformations
transform = Augmentations(
[
Resample(input_sr=48000, new_sr=16000, return_original_sr=False), # Downsample block
LFCC(
fs=16000, n_fft=512, hop_size=128, win_size=128, n_filters=128, n_lfcc=40 # LFCC augmentation
)
]
)
path_to_txt = '/home/adam/projects/audio_sleuth/audio_data/half_truth/HAD/HAD_train/HAD_train_label.txt'
# Returns torch Dataset that can be fed into DataLoader
dataset = HalfTruthDataset(
path_to_txt, # Initialize to your txt file of filepaths and labels from dataset
fs=16000
)
# You can feed this data_loader into your training
data_loader = DataLoader(dataset, batch_size=32, shuffle=True, collate_fn=lambda x: pad_and_transform_collate(x, transform))