| The speech recognition task can also be called Automatic Speech Recognition(ASR).At present,speech recognition technology has been integrated into all aspects of our life and work.These products containing speech technology reduce the learning cost of complex systems and improve our production efficiency and quality of life.However,with the popularization of education,movement of population,and advanced communication technology and Internet technology all promote international cultural exchanges,more and more multi-language mixed use has appeared on various occasions,making code-switching speech recognition has become one of the problems that need to be solved in the current speech recognition field.The main problem in the study of code-switching speech recognition in Chinese and English is sparse data and flexible code-switching.If the amount of data is sufficient,it can be done in accordance with the construction of a monolingual speech recognition system.However,Chinese-English code-switching speech recognition often encounters the problem of data sparseness.In order to solve the problem of data sparseness,you can consider learning with Chinese and English monolingual data,as well as transfer learning,semi-supervised learning and other methods.Chinese and English monolingual data usually have relatively rich vocabulary and pronunciation phenomena,so the monolingual data can be used to train the model to provide a better set of initial parameters for training the code-switching speech recognition model.In addition,the ChineseEnglish code-switching data has a certain correlation with the Chinese and English monolingual data.This article attempts to filter the monolingual data through data selection,and select effective and highly relevant data to expand the training set.The experiment proves Its effectiveness.This method may cause the model to be biased in a certain language,so balance is more important when adding single-language data.In addition,the code-switching pattern is studied,and the importance of data matching is proposed.When building a robust speech recognition system,data augmentation methods are often used to expand the training data.In addition to the more common data augmentation methods such as speed disturbance,noise addition,reverberation,and Spec Augment,speech synthesis can also be used as a data augmentation method,using text other than the data set to synthesize speech data,and adding training sets to alleviate data sparseness The problem.In addition,in low-resource data sets,there is often an imbalance in the amount of speaker data,so that the trained model does not have high recognition performance for speakers of a certain accent or age.There are certain speakers in the training set and test set.Does not match.In order to alleviate this phenomenon,this article attempts to build a multi-speaker speech conversion system,expand the training set data,increase the number of speakers,and expand the training data.Experiments prove the effectiveness of this method in low-resource scenarios. |
