Prosodic focus marking for voice-AI and human addressees

Abstract

Prosodic prominence (through increased pitch, intensity, duration) is thought to guide listeners’ attention to new information, which relies on the assumption that a rational listener will benefit from these prosodic cues. This study investigates production and perception of prosodic focus marking toward two types of addresses: a human and a voice assistant interlocutor, a potentially less-than-rational listener. Stimuli consisted of question-answer pairs, where American English speakers read identical sentences (e.g., “Jude saw the sun”) in response to interlocutors’ questions probing different foci (e.g., “Who saw the sun?”). Experiment 1 reveals consistent acoustic adjustments to mark focus on either the subject or object of a sentence: speakers increase vowel intensity and duration. In Experiment 2, we find that listeners reliably infer the intended information structure based on these acoustic adjustments. Across both experiments, we see no consistent difference in focus marking by type of interlocutor (human vs. voice assistant). However, listeners associate particular features (e.g., slower speech rate) with speech directed at voice assistants. Taken together, our findings suggest that while speakers apply communicative strategies from human-human interaction when addressing voice assistants, listeners expect a device-specific register.