Datasets:
Problem with splitting dataset for training for deepfake detection
Hi,
I would like to download this subset of the AUDETER dataset, but I’m not sure how to properly split it, since all the audios are generated.
Would it make sense to download the mls section of AUDETER dataset using load_dataset, take the dev and test parts from the original mls corpus mentioned in the paper, then mix, label them manually to real/spoof and split it to tran, dev, test?
It feels like it would be much more convenient if a properly pre-split version of the dataset was available — for example with balanced real vs. fake audio and three labeled splits (train, dev, test) ready to download.
Hi, do you know how to split it now?
Yeah, I am currently working on a project that uses this dataset, you can check here https://github.com/mkarapka/audio-deepfake-detection-uwr. I wrote a whole pipeline that separately downloads and preprocesses audio files from this dataset and mls_eng (with corresponding bonafide audio samples). There is a script that runs the whole pipeline - scripts/preprocessing/run_full_preprocessing_pipeline.py which outputs an .npy file with preprocessed features and a .csv file with labels for each audio. To split the data, there is another script that splits the .csv file into train, dev, and test sets - scripts/preprocessing/run_split_dataset.py. And to extract features for each split, I wrote a class called FeatureLoader that returns the corresponding features from the .npy file for each .csv split.