Research On Singer Recognition Key Technology

With the development of digital technology and networked multimedia,the scale of the digital music industry on the Internet continues to expand.How to effectively organize,browse and retrieve large-scale music content has become a key issue in the field of music information retrieval.Singer recognition uses modern computer technology methods to analyze a singer’s singing voice and extract the singer’s audio features,thus enabling the verification and identification of the singer’s identity.Based on deep learning techniques and big data technology,singer recognition is of great importance for music copyright protection and the development of digital music technology.Conventional datasets for singer identification contain background noise interference,thus building a dataset of a cappella songs is crucial for studying the acoustic information of a cappella singing and lyric recitation.In addition,what kind of sound source separation method is used for music datasets that contain background noise has become an urgent research problem.The paper begins with a detailed description of the relevant content within the field of music information retrieval,an introduction to existing singer recognition methods,and a brief illustration on the composition of music,including pitch,timbre and intensity.The singer recognition system is then studied and the acoustic features used are outlined,followed by a discussion of the theoretical knowledge of neural networks and deep learning,and an in-depth study of the development of deep learning and structural models.In this paper,model parameters designed for ECAPA-TDNN,X-vector and Res Net50 are applied to feature extraction and model training tasks for singer recognition systems for the first time.Then feature vectors are classified and identified by probabilistic linear discriminant analysis or cosine similarity discriminant.In the first part of this thesis,the experimental study of singer recognition is carried out on the a cappella singing dataset.A high-quality a cappella singing dataset is recorded and collected,a singer recognition system framework based on deep neural network is built,and feature extraction and singer model training are carried out using the a cappella singing data and the speech data respectively.The result shows that a cappella songs have better generalization performance due to their wider sound range and more complex features.The second part of the experiments proposes a singer recognition method based on source separation for Chinese and English network song datasets with background accompaniment,and applies two source separation models trained based on neural network feature mapping to carry out singer recognition experiments based on source separation in two stages: test stage and training test,respectively.The results show that the source separation method proposed in this paper has important significance and value for singer recognition,which can effectively eliminate interference factors in mixed signals and highlight vocal features,thus improving the accuracy and robustness of the singer recognition system.Taken together,the results of this paper demonstrates that a cappella singing data and source separation techniques have great research value in the field of deep learning-based singer recognition.