Training PASE architecture for detecting fake audios

[Zain-Jiang]

Hi, Thanks again for this great contributions! I had read the codes and paper. Using self-supervised learning in this task is really amazing. When I want to train this model for detecting fake audios using ASVspoof2019 LA dataset, I find some problems. There are countless kinds of fake audios from different people. If I simply using 1 for fake audios and 0 for bonafide ones, Is it appropriate? or if there might be some appropriate ways for me to handle this problem ?

labels=load_label()
dic=dict()

for item in labels:
    item=item.split(" ")
    if item[-1]=="bonafide":
        dic[item[0]+".flac"]=0
    else:
        dic[item[0]+".flac"]=1
np.save("ASVspoof2019_dict.npy",dic)