Shruti Kshirsagar

Speech enhancement can paradoxically *hurt* audio deepfake detection, as the algorithm that improved perceptual speech quality the most actually *reduced* the detection accuracy.

RAG systems readily absorb and amplify ideological biases present in retrieved documents, even more so when prompts explicitly describe the ideological dimensions at play.

Forget training on entire utterances: PhonemeDF reveals that phoneme-level analysis and KLD can pinpoint the most vulnerable spots in synthetic speech, boosting deepfake detection.