| With the development of information technology and artificial intelligence,multilingual speech human-computer interaction(HCI)systems have gradually developed in order to facilitate human life.Language identification(LID)is the front-end processing of multilingual speech HCI systems.Its effect of identification will directly affect the use of subsequent HCI systems.It also has great research significance and application value.How to improve the identification capability of LID system in complex sound scenes is a challenging problem in the field of LID at present.Therefore,it is necessary to further investigate embedding features that are independent of the environment and can represent language features.Based on the above analysis,the dissertation focuses on three issues:(1)how to make neural networks pay more attention to the frame-level features that work for the classification task;(2)how to capture the complex geometric distribution of speech data containing multiple local clusters;(3)how to measure the degree of association between features more fairly in deep metric learning.Ultimately,it will achieve the goal of improving the performance of the LID system.The main contents of the dissertation are summarized as follows:(1)In the aspect of making the network model more focused on the frame-level features that play an important role in the classification task,a LID method based on an extended delay neural network(E-TDNN)and multi-head self-attention is proposed.In order to give more weight to the frame-level features that play a greater role in the LID task and reduces the weight of other useless frame-level features when aggregating frame-level features,multi-head self-attention mechanism is introduced into E-TDNN in this method.The proposed method reduces the impact of distortion factors such as noise on the features and ultimately enhances the robustness of the embedding features.Meanwhile,the multiple heads in the self-attention mechanism can capture different linguistic discriminative information in the frame-level features.Finally,the discrimination of the embedding features is further improved.Experimental results show that the proposed method captures more discriminative embedding features than other methods,which effectively improves the performance of the LID system.(2)In the aspect of capturing the complex geometric distribution of speech data containing multiple local clusters,a LID method based on multi-center softmax is proposed.This method introduces multiple centers for softmax.By guiding the training of the network model,the method allows the features in each class to be approximately represented for each class by weighted summation over all centers of the class.It not only enhances the compactness within classes,but also can be more accurately to describe the complex distribution of the embedding feature space.Ultimately,the extracted embedding features are closer to the real class entities.Experimental results show that compared with the single-center softmax and other typical methods,the within-class features of the proposed method will be more compact.It can effectively mitigate the impact of complex sound environments on LID and improve the performance of LID system.(3)In the aspect of measuring the degree of association more fairly between features in deep metric learning,a LID method based on adjusted cosine mutual information estimation is proposed.The mutual information which cannot be susceptible to other factors is introduced into the objective function.It will maximize the mutual information of within-class features.Meanwhile,the within-class features will be more compact,and the between-class features will be more scattered.Then,the adjusted cosine similarity is introduced into the mutual information objective function.It can guide the neural network to use the common information of features in the same set to correct the similarity between features,so that the embedding features will be more robust.Experimental results show that compared with other methods,the proposed method can substantially improve the performance of the LID system. |
