An Adaptive ResNet Based Speaker Recognition in Radio Communication

In this paper, a speaker recognition strategy in military radio communication is applied. In military operations, the most commonly used method of information transmission is radio communication. Speaker recognition technology can confirm the sender's identity, and effectively prevent the enemy from pretending to be our military commander to issue false orders. However, the datasets of the military commander from the radio are confidential, and there are no large open-source datasets. Consequently, speaker recognition accuracy is not ideal if we only train a small sample of speaker datasets. Therefore, we propose a transfer learning method for training. We pre-train a Deep Residual neural network (ResNet) with large sample datasets and re-train a novel adaptive model with a simple sample dataset in radio communication. Experiments are carried out using the aishell-2 dataset and the self-collected radio military command datasets. Experimental results demonstrate that the adaptive network with transfer learning method improves the performance by 23.55% relatively compared to the baseline system in radio communication.