Improved Speech Authenticity Detection in Chinese-English Bilingual Contexts

Area of Science:

• Artificial Intelligence
• Speech Processing
• Cybersecurity

Background:

• The proliferation of voice technology necessitates advanced methods for authenticating speech.
• Existing audio tampering detection systems often lack robustness across different languages and tampering types.
• Spoofing attacks pose a significant threat to voice-based security systems.

Purpose of the Study:

• To develop and evaluate an improved model for detecting tampered audio, specifically addressing challenges in multilingual settings.
• To enhance the generalization capabilities of deep learning models for audio tampering detection.
• To benchmark the proposed model against state-of-the-art methods in recent competitions.

Main Methods:

• Developed a hybrid deep learning model integrating an enhanced ResNet architecture with a Long Short-Term Memory (LSTM) network.
• Created a bilingual dataset combining self-recorded Chinese speech and public English audio samples (VCTK2).
• Evaluated the model using advanced tampering techniques like CycleGAN voice conversion and auto splicing.

Main Results:

• The proposed ResNet-LSTM model achieved a superior performance with an equal error rate (EER) of 11.62% on a bilingual dataset.
• The model demonstrated superior performance compared to leading approaches from the ASVSpoof 2021 and ADD 2022 competitions.
• Effectiveness was validated against realistic tampering scenarios, including voice conversion and audio splicing.

Conclusions:

• The integrated ResNet-LSTM model offers a robust solution for detecting tampered audio, particularly in challenging multilingual contexts.
• The approach significantly advances the state-of-the-art in anti-spoofing technology.
• This work provides a foundation for more secure voice communication systems.