Deploying Emotion-Aware Speech GANs on Edge Devices: A Real-Time Optimization Framework

This study proposes a lightweight, real-time deployment framework for emotion-aware speech synthesis using Generative Adversarial Networks (GANs) tailored for edge devices. By integrating model optimization strategies such as structured pruning, post-training quantization, and ONNX-based cross-platform conversion, the framework enables efficient on-device inference across diverse hardware-including CPUs, GPUs, and the NVIDIA Jetson Nano. Performance evaluations highlight significant improvements, with inference latency reduced by up to $3.5 \times$ and memory usage by $2.8 \times$ on edge platforms, while retaining over $92 \%$ of perceptual quality based on Mean Opinion Score (MOS) metrics. On high-performance systems, the framework supports sub-50 ms latency, ensuring real-time responsiveness. These results affirm the feasibility of deploying expressive, emotionally adaptive speech systems in constrained environments. The proposed approach enhances user privacy through local processing and supports emerging applications in affective computing, conversational agents, and emotion-aware IoT systems. This work establishes a scalable benchmark for deploying generative audio models where both responsiveness and emotional nuance are vital.