An Improved Similarity Calculating Method for Automatic Speech Recognition

The paper presents a new calculation similarity method. It's called the constrained similarity method. Compared with the traditional cosine similarity and dot product method, the proposed method can maintain the similarity for proportional transformation vectors and also constrain the similarity result. In the attention mechanism encoder-decoder structure, the proposed method is applied. The proposed method obtains further improvement to the recognition results. The proposed method is more suitable for Wenet speech dataset which is closer to the real environment. The proposed method is more important practical significance.