Talk To Me: Testing Trust and Usability of Voice-Cloned Conversational Agents

Sammanfattning

Introduction: This study examines whether a conversational agent using a cloned version of the users’ own voice affects trust and usability compared to a standard synthesized voice. While prior research suggests that voice characteristics and perceived similarity may influence trust, it remains unclear whether cloned self-voices improve user trust, usability, or task performance. Research Question: In what ways are users’ trust and usability in a conversational user interface affected when the conversational agent uses cloned versions of users’ own voice compared to standard voices? Method: A quantitative between-subjects laboratory experiment was conducted with twelve participants. Participants completed a simulated flight booking task using a voice-based conversational agent. The control group interacted with a standard gender-matched voice, while the experimental group interacted with a cloned version of their own voice. Behavioral trust was measured through six predefined decision points during the booking process. Perceived trust and usability were measured using the Trust in Automation scale and the System Usability Scale. Task completion time, error frequency, recall performance, and objective voice similarity were also analyzed. Results: The results showed no statistically significant differences between the cloned-voice and standard-voice conditions for behavioral trust, perceived trust, usability, task efficiency, interaction errors, or recall. However, while results confirmed that the voice cloning manipulation produced significantly higher acoustic similarity in the experimental condition, subjective recognition was inconsistent. Only three of six participants in the experimental condition recognized the agent’s voice as their own. Discussion: The main contribution of the study is the distinction between objective acoustic similarity and subjective recognition of a cloned voice. Although cloned voices may be technically similar to users’ own voices, this does not necessarily mean that users experience them as self-like. In addition, exploratory open-ended responses suggested that participants’ trust was shaped by the visibility and verifiability of the system’s actions. The findings suggest that voice cloning may have limited standalone impact on trust and usability in short task-oriented interactions. Due to the small sample size, the results should be interpreted as exploratory.