BPCNN: Bi-Point Input for Convolutional Neural Networks in Speaker Spoofing Detection.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Signal Processing

Background:

• Convolutional Neural Networks (CNNs) typically require fixed-size inputs, posing challenges for variable-length data like speech.
• Existing methods like feature segmentation limit CNNs to processing one segment at a time, hindering comprehensive analysis.

Purpose of the Study:

• To propose a novel input method, "bi-point input," for CNNs to effectively handle variable-length features.
• To enhance the information processing capacity of CNNs by allowing them to consider multiple segments simultaneously.

Main Methods:

• The proposed bi-point input method feeds pairs of segments from a variable-length feature into a CNN concurrently.
• Various combination strategies for these segments are explored, with guidance provided for optimal segment length selection.
• The method was evaluated on spoofing detection tasks using the ASVspoof 2019 database.

Main Results:

• The bi-point input method significantly improved performance in spoofing detection tasks.
• A relative reduction in equal error rate (EER) of approximately 17.2% for logical access (LA) and 43.8% for physical access (PA) was achieved.

Conclusions:

• The bi-point input method offers a more effective way for CNNs to process variable-length features compared to traditional segmentation.
• This approach enhances the contextual information available to the model, leading to improved accuracy in tasks like speaker verification and anti-spoofing.