Abstract
Abstract This paper introduces a novel similarity calculation method, called unit dot product similarity method. The proposed method restricts the denominator by the sum of vector modulus. Compared with traditional dot product similarity calculation method, the proposed method can maintain the similarity of equally scaled vector and obtain the bounded similarity result. We develop and compare the proposed method in the attention-based encoder decoder structure. The proposed method brings the recognition results further improvement. For the end-to-end speech recognition model, we select greedy soup instead of the average model parameters in WeNet. A dynamic parallel greedy soup optimization algorithm is proposed to increase computational speed. The experiments show the importance of proposed method and optimization algorithm. The effectiveness is also proved on multiple corpora.