Introduction
Traditional Conversational Recommender Systems (CRSs) are predominantly text-based (TCRSs). While effective, TCRSs present significant barriers for users with visual impairments, limited literacy, or those in "eyes-busy" environments like driving. Furthermore, text interfaces often suffer from sparse or inaccurate user profile data. This paper explores the transition to Voice-based Conversational Recommender Systems (VCRSs), arguing that voice is a more natural, efficient, and "honest" modality. A user's voice inherently carries metadata—such as age and gender—that is difficult to falsify and highly valuable for personalization.
Methodology
The authors propose a multi-stage pipeline to transform static interaction data into a rich voice-based conversational environment. This pipeline bridges the gap between raw user-item interactions and natural vocal dialogue.
The process begins with Backbone Selection, utilizing datasets like Coat (e-commerce) and MovieLens-1M. ChatGPT-Driven Dialogue generation then creates natural-sounding conversation templates, moving away from rigid robotic prompts. These dialogues are converted to audio using VITS, a state-of-the-art neural speech synthesis model, with speakers assigned from the Vctk dataset based on verified user metadata. Finally, a Multi-task Classification Module (MCM) extracts demographic features from the audio to inform a Voice Feature Integration Module (VFIM) for final recommendations.
Dataset Benchmarks
The study introduces the first public benchmarks for VCRS: VCRS-Coat and VCRS-ML-1M. Unlike previous datasets like ReDial or OpenDialKG, these include synthesized voice audio aligned with verified user metadata, allowing for the training of models that "hear" user characteristics.
Key Insights and Results
A core finding is the "honesty" of voice. While users may skip demographic fields in forms, their vocal acoustics provide reliable signals. Interestingly, the study found that while humans excel at gender identification (95%+), they struggle with age detection (30-50% accuracy). The proposed MCM model significantly outperformed humans in age detection, identifying vocal patterns invisible to the human ear.
Empirical results demonstrate that side information is critical: removing age and gender features led to a 16.3% drop in recommendation performance (NDCG@10) on the Coat dataset. Furthermore, the synthesized audio achieved a high Mean Opinion Score (MOS) of 4.73/5.0, proving its viability for real-world application.
Conclusion and Future Work
This work lays the foundation for VCRSs by providing public datasets and a robust feature-fusion framework. Future research should address the limitations of simulated conversations by incorporating real-world human recordings and exploring end-to-end dynamic systems trained via reinforcement learning to handle environmental noise and multi-speaker scenarios.
Summary updated