Abstract: In order to ensure the accuracy and real-time of speech recognition, a speech recognition method based on time synchronization recursive attention mechanism is proposed. The windowless attention mechanism was introduced which did not require multiple training sessions to save model preparation time, and the context vector was obtained by using the time synchronization recursive update rule instead of the formula based on the kernel function smoother. The tradeoff between delay and performance was further controlled by adjusting the scalar threshold related to the attention endpoint decision. Experiments show that the proposed method can not only ensure the recognition accuracy, but also achieve online recognition.