Spatial reconstructed local attention Res2Net with F0 subband for fake speech detection

Area of Science:

• Speech Processing
• Artificial Intelligence
• Signal Analysis

Background:

• Synthetic speech often lacks the natural rhythm and fundamental frequency (F0) variations of genuine human speech.
• The fundamental frequency (F0) is a crucial feature for distinguishing between real and synthetic speech.
• Existing fake speech detection (FSD) methods may not fully leverage the discriminative potential of F0 variations.

Purpose of the Study:

• To propose a novel method for fake speech detection (FSD) by focusing on the fundamental frequency (F0) subband.
• To enhance the modeling of F0 subbands for improved FSD performance.
• To achieve state-of-the-art results in fake speech detection.

Main Methods:

• A novel F0 subband approach is proposed for fake speech detection (FSD).
• The spatial reconstructed local attention Res2Net (SR-LA Res2Net) architecture is introduced to effectively model F0 subbands.
• Res2Net backbone extracts multiscale information, enhanced by spatial reconstruction to preserve channel data and local attention to focus on F0 subband details.

Main Results:

• The proposed method achieved an equal error rate (EER) of 0.47% on the ASVspoof 2019 LA dataset.
• The minimum tandem detection cost function (min t-DCF) reached 0.0159.
• The system demonstrated state-of-the-art performance compared to other single systems.

Conclusions:

• The proposed F0 subband method, utilizing SR-LA Res2Net, is highly effective for fake speech detection (FSD).
• The integration of spatial reconstruction and local attention significantly enhances the modeling of F0 features.
• The achieved results represent a significant advancement in the field of fake speech detection.