AI Voice Detector

• The system focuses on comparing metric outputs from Handcraft Feature, Transformer models, and attempting to enhance Rawnet2 with them.

Dataset

English

• ASVspoof 2021 (Large-scale dataset used for AI voice detection competitions, containing both genuine and spoofed voices, total 56251 files or 7.23 GB)

Thai

• Mozilla Common Voice Corpus 16.1 First Validated 10000 files (Bonafide)
• AI For Thai Fake audio for thai voice 49891 files (Spoof)

Front-end

• Raw

• Handcrafted Feature

  • LFCC
  • Mel spectrogram

• Feature Extractor

  • Wav2vec

Classifier

• RawNet2
• Random Forest
• ANN (Artificial Neural Networks)
• KAN (Kolmogorov-Arnold Networks)

Metric

• Accuracy
• min t-DCF (tandem detection cost function)
• EER (Equal Error Rate)

Our Experiment

• 

Deploy on

• pythonanywhere by ANACONDA

Framework

• Django

Check our website here 🚀🚀🚀

• https://vocalsense.pythonanywhere.com/