search

josebeo2016 / BTS-Encoder-ASVspoof

Synthesis speech detection based on Breathing-Talking-Silence sounds
16 stars 1 forks source link

Training from scratch #8

Open magdalenapasternak opened 1 week ago

magdalenapasternak commented 1 week ago

Hello.

It looks like the training that you provided uses CNNClassifier and has been retrained in biosegment folder: In predict.py: classifier = CNNClassifier(os.path.join(BASE_DIR, "out_datanewjun30", "cnn.pth"), config, device=device)

In order to train it, we need: segments, utt2spk, and wav.scp files.

The files you have provided in the bio segment folder seem not to include all the ASVspoof training/dev folders. Can you share how to either create these files or share whole files?

josebeo2016 commented 1 week ago

The CNNClassifier for Sound Segmentation is our extended work (not in the ICASSP 2023 paper). You can change it to GMMClassifier to reproduce this work. The Sound Segmentation module is the first stage of the system. In biosegment folder, we provided the checkpoint, so that you can use it directly, or you can train it again following this discussion: https://github.com/josebeo2016/biosegment/issues/6