The paper is planning to the challenge of the speaker's acquisition (ASD) when the program requires determining at the real time the person is speaking or not in the series of video frames. While previous jobs have made important enhancements in improving network buildings and reading practical Asd presentations, the sensitive gap is in the original-time submission gap. Existing models are always suffering from high latency and memory usage, contributing to immediate applications. Closing the gap, we present two situations that discuss important challenges caused by real-time issues. First, we introduce how to reduce the number of future related frames used by the ASD model. In doing so, we reduce the need for processing all the order in the future frames before a decision is made, highly reduces the latency. Second, we suggest that there are strong problems that limit the total number of past model that can reach the step. This deals with persistent memories associated with Asd broadcasting programs. In addition to these edition structures, we conduct comprehensive trials to ensure our method. Our results indicate that transformer models are pressed can carry out comparisons comparable or better than high quality models, such as Uni-Direct Grus, with a highly reduced number of frames. In addition, we illuminate temporary memory requirements for Asd Systems, indicating that the main context of the past has a major impact on accuracy than the accuracy. When confirming the CPU we find that our active Buildings are a memory fastened with the number of previous back condition we can also use that the computer cost cannot be comparable to memory costs.
