Attractive synthetic voices

With recent advances in Artificial Intelligence (AI), synthetic voices have become increasingly prevalent in our everyday soundscape. This study examined listeners’ perception of human and neural Text-To-Speech (TTS) voices. In an online experiment, 75 participants listened to different versions of a short utterance spoken by eight different voices (half human, half TTS), each presented in four expressed emotions (neutral, happy, sad, angry). For each stimulus, participants rated voice attractiveness and willingness to interact, and selected the perceived emotion from a forced-choice list. In a second part, participants were asked to classify each voice as human or AI-generated. Results revealed that participants were often “fooled” by the TTS voices, misidentifying them as human. Voice ratings were influenced by the perceived emotion regardless of the voice type, with happy-sounding voices rated more positively than those perceived as sad or angry. However, TTS voices were rated as less attractive and socially appealing overall, though with large individual differences. These findings indicate that TTS voices are approaching human ones in how they are perceived by listeners, highlighting progress in their naturalness. • Human voices are still rated as more attractive than synthetic ones – but perhaps not for long. • Participants are often fooled by TTS voices, misidentifying them as human. • Perceived emotion influences ratings similarly for human and TTS voices, with strong effect of valence. • Older adults are less accurate in distinguishing TTS from human voices.