Framing, more than speech, affects how machine agents are perceived

We tested how speech phenomena (discourse markers like oh and you know and fillers like uh and um) and the conceptualisation of voices (as human or machine) affected speaker quality ratings (trustworthiness, friendliness, intelligence, and nervousness) across dialogues and monologues. Listeners preferred human voices over synthesised voices in dialogue containing discourse markers and fillers (Study 1). Shifting the location of discourse markers to new locations did not negatively impact ratings of what participants heard (Study 2), but hearing shifted markers did lead participants to rate imagined machine speech containing markers more negatively on a post-experiment questionnaire. Backstories also impacted perceptions of recorded speech: People rated the same synthesised speech more positively if they believed it was created by a person than if they believed it was created by machine (Study 3). Similarly, people's opinions about robots varied depending on context (Study 4). How people rate machine agents in the future will likely reflect how machine agents are portrayed to the public, capabilities aside.