TRU-Net Based AI Model Implementation by Pure C for Real-time Speech Enhancement

This paper presents a real-time speech enhancement AI model based on TRU-Net, implemented in Pure C for deployment on resource constrained edge devices. High-level machine learning frameworks such as PyTorch and TensorFlow can impose significant limitations in edge environments due to their memory footprint, power consumption, and real-time control constraints. Accordingly, this study emphasizes the importance of architecture design that prioritizes efficient memory utilization in addition to computational performance optimization. By implementing the model in Pure C, memory access patterns, buffer sizes, computational methods, and parallel processing structures are precisely controlled, enabling an experimental analysis of the trade-offs between performance and memory usage. Based on these analyses, an optimized NPU architecture is proposed that minimizes memory consumption, improves the efficiency of intermediate tensor and buffer management, and enhances parallel processing performance through loop unrolling. Experimental results demonstrate that the proposed Pure-C-based model significantly reduces per-frame memory usage and achieves more than a 70% reduction in execution time, validating its effectiveness for real-time, low-power edge AI systems.