Building LLM-Based Voice Agents for Requirements Elicitation: An Experience Report on Early Prototypes

Abstract

<p>This paper presents an experience report on the development and pre-testing of voice-based agentic workflows that uses two large language models, OpenAI’s GPT-4o-mini and Google’s Gemma3:27b to conduct requirement elicitation discussions in software projects. The growing use of independent AI agents for specialized tasks has motivated our exploration of voice agents as "requirements elicitors" within software projects. The paper describes the approaches attempted during development, including those that were successful and those that failed, along with insights gathered from implementing and testing the use cases. We conducted a pre-test round with five participants, comparing the performance of the two agents under two case studies. At this stage, the OpenAI-based agent showed a higher requirements coverage, identifying 77.5% of relevant requirements on average, while the Gemma-based agent captured 35.0%. In terms of usability, participants rated the OpenAI agent 4.0/5, compared to 3.3/5 for the Gemma agent, highlighting a more natural conversational flow, better contextual understanding, and improved responsiveness. We propose deploying this voice agent as the first agent in an extended multi-agent requirements engineering workflow to support requirement elicitation sessions alongside a human requirement engineer, which will be the extension of this work. The methods, design choices, and lessons learned documented in this report aim to guide practitioners and researchers in adapting similar agent-based approaches in their own domains.<br></p>