Low-Complexity Voice Activity Detection Algorithm for Edge-Level Device

This paper presents two optimization techniques to relieve the computational complexity of the neural network based voice activity detection (VAD) task. Proposed techniques analyze the similarity between speech features by comparing the vectors at adjacent time steps and reduce the required computational cost by modifying internal elements based on the similarity. As a case study, a simple convolutional neural network for VAD was simulated with the proposed optimization techniques under the noisy environment, and experimental results show that the proposed techniques can reduce the required computational cost up to 33.6% with negligible performance degradation.