Speech deepfake detectors are usually tested on utterances with a consistent emotional tone, but modern expressive synthesis can change emotion mid-sentence. A new paper on arXiv introduces EmoShift, a benchmark built to probe this gap.
The benchmark compares emotion-stable (AA) utterances with emotion-changing (AB) utterances. The authors use this setup to expose failure modes in current detectors, showing that intra-utterance emotion shifts are a weak spot that standard evaluation sets miss.
As synthetic voices become more expressive, detection systems will need to account for emotional variation within a single utterance rather than assuming a flat, uniform delivery.