ShiftySpeech is a large-scale synthetic speech detection benchmark released by Johns Hopkins University in 2025. The related paper is titled "ShiftySpeech: A Large-Scale Synthetic Speech Dataset with Distribution ShiftsThe aim is to study the generalization ability of speech synthesis detection models in the real world when faced with "distribution drift" (including changes in language, speaker, generation model, and recording conditions).

This dataset contains over 3,000 hours of synthesized speech, covering seven source domains, including reading styles, podcasts, YouTube recordings, and other scenarios with background noise or non-standard recording conditions, as well as variations in language, speaker age, accent, and gender. The data covers three languages (English, Chinese, and Japanese), and speech was generated using six TTS (text-to-speech) systems and twelve vocoders (vocoders/waveform generators) to construct different degrees of system distribution drift.