Cross-Technology Generalization in Synthesized Speech Detection: Evaluating AST Models with Modern Voice Generators

TL;DR

The paper tackles the problem of detecting synthesized speech and generalizing detection across emerging voice generators. It evaluates the Audio Spectrogram Transformer (AST) with differentiated augmentation to distinguish synthetic artifacts from bona fide speech, demonstrating rapid adaptation with as few as 102 training samples. The results show an overall EER of 0.91% across tested technologies and an average of 3.30% on unseen generators when trained on ElevenLabs data alone, indicating cross-technology generalization. This work provides benchmarks for rapid adaptation and supports transformer-based approaches as robust tools for real-world speech verification amid evolving synthesis technologies.

Abstract

This paper evaluates the Audio Spectrogram Transformer (AST) architecture for synthesized speech detection, with focus on generalization across modern voice generation technologies. Using differentiated augmentation strategies, the model achieves 0.91% EER overall when tested against ElevenLabs, NotebookLM, and Minimax AI voice generators. Notably, after training with only 102 samples from a single technology, the model demonstrates strong cross-technology generalization, achieving 3.3% EER on completely unseen voice generators. This work establishes benchmarks for rapid adaptation to emerging synthesis technologies and provides evidence that transformer-based architectures can identify common artifacts across different neural voice synthesis methods, contributing to more robust speech verification systems.

Figures (4)

• Figure :
• Figure :
• Figure :
• Figure :