Establishing design principles for voice-enabled conversational agents serving the reflective practice

Research context

Professional education increasingly requires students to develop metacognitive skills through reflective practice, yet students consistently fail to reach critical reflection levels and remain descriptive. Students perceive reflective writing as time-consuming with unclear benefits and consistently express preference for reflecting in conversations instead of writing. Recent conversational agents (CAs) support reflective writing, but remain text-focused despite students' stated preference for voice interaction and recent technological advances making voice interaction viable. This thesis addresses a current research gap: no systematic design principles exist for voice-enabled conversational agents serving reflective learning in higher education.

Research question and objectives

How can a conversational agent for reflective purposes be designed with voice capabilities as a viable input method for university students?

This thesis followed the Design Science Research Methodology (Peffers et al. 2007) to:

• Systematically identify requirements from users and literature
• Develop empirically grounded design principles for voice-enabled reflective CAs
• Operationalize these principles in a working prototype (VoxaReflect)
• Validate and refine the principles through authentic user testing

Methodology

The research combined a literature review across reflective learning, CA design, and anthropomorphism and semi-structured interviews with 9 university students to form initial design principles. The prototype VoxaReflect was built based on these design principles and tested with 6 participants conducting authentic reflection tasks. In a last step, the design principles were refined and used for an iteration of VoxaReflect.

Key findings

The main contribution are the seven design principles that are formulated using Gregor et al.'s (2020) schema for transferability:

1. Voice-centred, chunked interaction – Enable voice as primary modality with text fallback, reliable ASR, and clear visual feedback
2. Phased dialogue – Guide users through structured reflection models (Gibbs' cycle) with phase tracking and progression safeguards
3. Ensure depth – Move students from description to critical reflection through open-ended questioning
4. Preserve engagement and ownership – Scaffold through questions rather than content generation
5. Short sessions with progress and payoff – Design for repeated practice with clear indicators and summaries
6. Conversational styles and expression – Offer selectable interaction styles while maintaining technical self-description
7. Transparent anthropomorphism – Communicate artificial nature explicitly despite voice cues

Validation results
Testers reported a strong preference for VoxaReflect over writing-based systems while rating the system well in TAM (Technology Acceptance Model) scores. They also reported equal or better reflection depth compared to written methods while remaining aware of the AI's nature despite anthropomorphic voice.

Critical success factors
Testing revealed that voice viability depends on technical reliability, optimized latency, UI responsiveness (visual feedback for recording and progression) and a flexible session design (user control over duration and depth).

Contribution

This thesis provides:

• First systematic design principles for voice-enabled reflective CAs in higher education
• Empirical validation that voice succeeds for structured reflection, contradicting previous findings from mobile expressive writing research
• A working prototype demonstrating feasibility with current AI technologies.

The seven design principles offer actionable guidance for researchers and developers building similar systems, showing that voice can lower barriers to reflective practice while maintaining pedagogical rigor, when technical reliability, ownership preservation and transparency are properly addressed.