A Thai Spoken Dialogue Generation System Incorporating Generative Adversarial Networks

Abstract

Traditional Thai dialogue generation systems, under maximum likelihood training, often favor high-frequency "safe" sentence structures, resulting in monotonous style and poor multi-turn contextual coherence. Similar problems exist in automated quality control within the textile industry: traditional models, trained on large amounts of defect-free data, struggle to distinguish subtle or rare defects, and are difficult to capture long-range dependencies in multi-sensor data streams. This paper proposes a Thai spoken dialogue generation system based on RelGAN and Transformer. It captures multi-turn context through a segmented relative position self-attention encoder and segmented recurrent memory, while the decoder embeds a relational memory module to track dialogue states and long-term entity slots, simultaneously generating text and prosodic markers. The generated results are prosodic mapped using FastSpeech2 and synthesized into high-fidelity speech using HiFi-GAN. Experiments show that the system achieves a Distinct-2 score of 0.11, an intent retention rate of 97.5%, an average response time of 98.44 ms, and a MOS of 4.05. This method has significant potential for cross-industry applications, enabling the construction of highly consistent real-time diagnostic systems in the textile industry, processing continuous multi-cycle sensor data, and achieving rapid fault identification and feedback.