Recent advances in Artificial Intelligence (AI) have led to highly realistic generative speech models, making it difficult to distinguish synthetic from natural speech. Technologies like Google Duplex, leveraging WaveNet technology, a deep neural network for seamless speech creation, exhibit an impressive degree of realism and naturalness. This highlights many opportunities such as realistic voiceovers and risks related to privacy and security. This work investigates methods from the ASVspoof challenge to develop a system capable of detecting synthetic speech in Portuguese from Portugal (PT-PT). To the best of our knowledge, this is the first such effort for this language. We introduce a PT-PT dataset with 22050 examples with both real and synthetic speech, the latter created using both text-to-speech and speech-to-speech systems. Several models from the literature were implemented and evaluated across English and PT-PT datasets. Statistical analysis identified ResNet18 and ECAPA-TDNN as top performers. After fine-tuning, ECAPA-TDNN achieved over 70% accuracy across all datasets, establishing itself as the most promising model.

错误:搜索内容不能为空,请输入英文关键词
错误:关键词超出字数限制,请精简
高级检索

PT-PT Synthetic Speech Detection

  • Rafael Geraldo dos Santos,
  • Joana Sousa,
  • José Valente de Oliveira

摘要

Recent advances in Artificial Intelligence (AI) have led to highly realistic generative speech models, making it difficult to distinguish synthetic from natural speech. Technologies like Google Duplex, leveraging WaveNet technology, a deep neural network for seamless speech creation, exhibit an impressive degree of realism and naturalness. This highlights many opportunities such as realistic voiceovers and risks related to privacy and security. This work investigates methods from the ASVspoof challenge to develop a system capable of detecting synthetic speech in Portuguese from Portugal (PT-PT). To the best of our knowledge, this is the first such effort for this language. We introduce a PT-PT dataset with 22050 examples with both real and synthetic speech, the latter created using both text-to-speech and speech-to-speech systems. Several models from the literature were implemented and evaluated across English and PT-PT datasets. Statistical analysis identified ResNet18 and ECAPA-TDNN as top performers. After fine-tuning, ECAPA-TDNN achieved over 70% accuracy across all datasets, establishing itself as the most promising model.