Evaluating trustworthiness across ethnically diverse human and commercial synthesised voices: A comparative study

This study examined trustworthiness perceptions in the tone of voice of human and real-world synthesised voices, focusing on the impact of acoustic features, speaker and listener ethnicities, listener biases toward voice-based intelligent agents and speaker nature (human vs synthesised). Speech rate, mean pitch, harmonics-to-noise ratio, jitter, shimmer, cepstral peak prominence, and long-term average spectrum, significantly influenced trustworthiness ratings across both human and synthesised voices. Synthesised voices were rated as sounding more trustworthy than human voices with no explicit intent behind their tone of voice (i.e., neutral). However, synthesised voices were rated as sounding less trustworthy than human voices when human speakers intentionally attempted to sound trustworthy. Moreover, listener biases were measured using the Negative Attitudes toward Robots Scale (NARS), where a general scepticism toward robots lowered trustworthiness ratings overall. White speakers were consistently rated as more trustworthy than Black or south Asian speakers across all listener ethnic groups. The findings highlight the need to optimise acoustic properties of synthesised voices for trustworthiness while addressing biases related to speaker ethnicity and listener attitudes toward robots. • Synthesised voices rated as more trustworthy than human voices with neutral intent. • Human voices with trustworthy intent outperformed real-world synthesised voices. • Synthesised acoustics hit a “perceptual middle ground” between human voice intents. • Listeners' negative attitudes toward robots and ethnicity affect trust ratings. • Study offers research-informed guidelines to design trustworthy synthesised voices.