Intelligent Virtual Assistant for Survey Research: From Development to Field Evaluation

This study presents the development and field evaluation of an Intelligent Virtual Assistant (IVA) designed to administer structured surveys in real-world settings. The system integrates a browser-based questionnaire, visual stimuli, respondent-selected avatars, speech-to-text and text-to-speech services, and a Large Language Model decision layer that interprets spoken responses, maps them to predefined categories, and provides clarification when needed. The IVA was deployed during the housing fair (2025) in Luxembourg, where 103 participants completed an English or French survey using either voice or click-based responses. Results demonstrate that the prototype was operationally feasible and generally well received. Respondents rated the assistant positively for functionality, conversation quality, privacy, and responsiveness, although confidence in the system's ability to understand users was comparatively weaker. Voice interaction was also selective: 61% of respondents answered at least three questions orally, while 23% used voice for all items, and greater discomfort with talking to computers was associated with lower voice uptake. System logs showed that most spoken turns were processed successfully, but a substantial minority required clarification or waiting. Performance was further affected by speech-recognition errors, ambient noise, accents, question type, and response format, while French sessions generated more discarded partial speech-recognition events. The public setting may also have discouraged some respondents from answering aloud because of privacy or social exposure concerns. The findings position IVA-administered surveys as a hybrid mode between self-administered and interviewer-led data collection. Their potential depends on mixed-modality design, robust field implementation, question-level optimization, and careful management of interaction failures and user trust.