Fast QBE: Towards Real-Time Spoken Term Detection with Separable Model

State-of-the-art spoken term detection or query-by-example networks depend on recurrent neural network (RNN), which extract fixed-dimensional vectors (embedded vectors) from both spoken query and the search content, and then calculate cosine distances over the vectors. However, these methods depend on time sequence, so it is a computational cost task, can not meet the requirements of both the query accuracy and search speed. In this work, we introduce a fast Spoken term detection system based on a separable model—RepVGG. Because of the trick of reparameterization, it has a faster speed in inference. Secondly, we use non maximum suppression and norm in the step of inference to improve it performance. Thirdly, we use multilanguage training to improve both accuracy and robustness of the system. Corresponding experiments are designed to verify these ideas. It show that proposed methods can import the GPU real-time factor (RTF) from 150 to 2300, and outperforms the state-of-the art method.