In the UCL study, researchers used a text-to-speech (TTS) algorithm trained on publicly available datasets in English and Mandarin to generate 50 deepfake speech samples in each language.