Enhancing Audio Anti‐Spoofing Through Innovative Utilization of Bona Fide Data, One‐Class Learning, and Generative Adversarial Networks

Name

Zihao Wei

Major

Data Science

Class

2023

About

Zihao Wei was born in 2001 in China. He is completing his bachelor degree in Duke Kunshan University with a major in Data Science.

Signature Work Project Overview

The author proposes a novel approach to enhance audio anti-spoofing using innovative techniques. Specifically, the use of “Golden Silence” in a partial spoofed dataset is explored and its advantage were took by transforming the binary classification (bona fide vs. spoofed) task to a triple-label (bona fide vs. spoofed vs. silence) one on a segmental level. Further delving into the bona fide data, the author proposes a one-class GAN approach with perturbation on embeddings to train a more robust and stronger discriminator tested on ASVspoof2019 dataset.

This work is is primarily focused on programming and entails a substantial amount of coding workload.

Signature Work Presentation Video

SIGNATURE WORKCONFERENCE & EXHIBITION 2023