Research on target speech extraction in similar language environment
Aiming at the common similar speech separation problem in practical applications,a new database,P-C,is con-structed to simulate similar language environments.This database combines the self-made Chongqing dialect dataset CQSpeech and the publicly available Chinese dataset THCH30 as a way to study the separation problem of mixed speech between Mandarin and Chongqing dialects.In addition,in order to fully utilize the speech features,speaker features are embedded in the CRN net-work.A large amount of data is first trained by the model to obtain speaker features,and then the speaker features are fused with the features in the separation model.This can effectively improve the clarity and accuracy of similar language speech separation.According to the experimental verification,the model demonstrates good separation effect on P-C database.
万方数据
维普
