Design and Evaluation of a Hybrid Georgian Voice Assistant for Public Services

Abstract

Voice assistants have become an essential component of modern human-computer interaction, allowing users to access information through natural speech. However, most commercial voice assistants provide limited support for low-resource languages. One such language is Georgian, for which limited support for voice technologies makes it difficult for citizens to access local public services through voice interaction. This limitation is especially important for elderly users and people with visual impairments, who often use voice interfaces instead of traditional graphical user interfaces.

This paper presents the design, implementation, and evaluation of a hybrid Georgian language voice assistant that provides voice access to information about public transport and cultural events across Tbilisi. The proposed system combines cloud-based speech recognition, natural language understanding, and speech synthesis technologies within a web application architecture. Speech recognition is performed using Google Cloud Speech-to-Text Chirp 3 [3], while speech synthesis uses Microsoft Edge Neural Text-to-Speech. Intent recognition combines rule-based processing with Google Gemini 2.5 Flash, creating a hybrid architecture that improves robustness while maintaining low latency. Real-time information is obtained from TKT.ge and Google Maps services.

The developed prototype was evaluated through technical performance measurements and user-centered usability testing. Experimental results show that the proposed hybrid architecture provides reliable Georgian speech processing with acceptable latency and high usability.