Edge AI를 위한 Pure-C 모델을 활용한 TRU-Net 기반 실시간 음성 개선

초록

Developing AI models using machine learning frameworks such as TensorFlow or PyTorch often introduces limitations in memory usage, power efficiency, and real-time control. For edge AI deployment, optimizing computational performance alone is insufficient; architectures must also be designed to maximize memory efficiency and resource utilization. In this paper, we implement a TRU-Net based ambient noise reduction based on real-time AI model entirely in pure C to enable fine grained detailed control of system resources, including memory access, buffer size configuration, computational strategies, and parallel processing. The proposed implementation incorporates several optimization techniques, including optimized NPU architecture, minimal memory usage for speech enhancement tasks, efficient buffer and intermediate tensor handling, and improved parallelism via loop unrolling. Experimental results show that the proposed implementation reduces CNN execution time by 40% while minimizing memory usage, demonstrating its effectiveness for real-time, low-power edge AI applications.

키워드